Local-first dynamic memory for apps, agents, notebooks, and tools.
WaveMind stores memories in SQLite, finds relevant candidates with vector search, then uses a wave-field priority layer to decide what still matters: hot facts rise, stale facts fade, temporary facts expire, and namespaces keep users or projects isolated.
Quick Start | CLI | Studio | Python Example | HTTP Example | Where Data Lives | LangChain | Chroma Migration | Use Cases | HTTP API | Benchmarks | Benchmark Brief | Research Branches | Roadmap | Contributing | Limitations
WaveMind is a dynamic memory engine you can embed in a product.
Use it when your app needs to remember things like user preferences, decisions, corrections, notes, research snippets, support history, agent context, or temporary facts.
The short version:
normal vector search: find the nearest text
WaveMind: find the nearest useful memory
WaveMind is not trying to replace every vector database. It is the memory layer around retrieval: persistence, namespaces, TTL, hotness, priority, decay, explicit forgetting, audit events, and optional graph dynamics.
| Question | Answer |
|---|---|
| What does it store? | Text memories, vectors, metadata, tags, TTL, priority, and recall state. |
| Where does it store data? | A local SQLite file by default; Postgres is available for production state. |
| How do I use it? | CLI, Python API, FastAPI HTTP server, LangChain memory, or framework adapters. |
| What is different from Chroma/Qdrant? | WaveMind adds memory policy: hotness, decay, TTL, correction handling, and scoped recall. |
| When should I not use it? | For huge static document search where a mature vector DB is already the right tool. |
| What is the simplest install? | python -m pip install wavemind |
| If you need... | WaveMind gives you... |
|---|---|
| Memory that survives restarts | One SQLite file stores text, vectors, metadata, TTL, and recall state. |
| Per-user or per-project recall | Namespaces and tags keep memories separated. |
| Temporary facts | ttl_seconds lets facts expire automatically. |
| Corrections and changing preferences | Newer or reinforced memories can outrank stale ones. |
| A simple integration path | Python API, CLI, FastAPI server, and LangChain memory class. |
| Production hygiene | Backups, audit log, API keys, rate limits, Prometheus metrics, and OpenTelemetry traces. |
The shortest path from install to first recall:
python -m pip install wavemind
wavemind remember "Andrey is a trader" --namespace demo
wavemind query "What does Andrey do?" --namespace demoNeed a reminder after install?
wavemind quickstartWant to see and manage memory in a browser?
wavemind studioBy default, WaveMind creates wavemind.sqlite3 in the current working
directory. That file is the local source of truth. Keep it out of git and back
it up like application state.
Start here if you only want to use WaveMind from the terminal:
| Goal | Command |
|---|---|
| Show first-run help | wavemind quickstart |
| Store a memory | wavemind remember "Andrey prefers short answers" --namespace user:42 |
| Search memory | wavemind query "answer style" --namespace user:42 |
| Consolidate active patterns | wavemind consolidate --namespace user:42 --seed "Rust compiler systems" |
| Open local dashboard | wavemind studio |
| See stored state | wavemind stats --namespace user:42 |
| Delete a namespace | wavemind forget --namespace user:42 |
| Import notes | wavemind import ./notes.txt --namespace project:alpha |
| Use another database file | wavemind --db ./state/memory.sqlite3 query "budget" --namespace user:42 |
| Start the HTTP API | wavemind --db ./state/memory.sqlite3 serve --host 127.0.0.1 --port 8000 |
After this point, choose the integration path you need: Python, HTTP, LangChain, framework adapters, benchmarks, or production deployment.
WaveMind Studio is the built-in local dashboard. It runs on top of the same FastAPI app and SQLite database as the CLI:
wavemind studioIt opens http://127.0.0.1:8000/studio and gives you:
| View | What it is for |
|---|---|
| Memory map | See field energy as a heatmap. |
| Namespace explorer | Inspect memories per user, project, agent, or tenant. |
| Live query tester | Test recall before wiring it into an app. |
| Feedback buttons | Mark recalled memories as useful or not useful. |
| Import/export | Import local files and export a namespace snapshot. |
| Backup | Create SQLite backups from the browser. |
| Conflict visualizer | Inspect correction groups when memories disagree. |
| Memory OS Insights | See read-only hot-query, policy, execution-plan, and architecture suggestions before running background workers. |
For a server-safe local bind:
wavemind --db ./state/wavemind.sqlite3 studio --host 127.0.0.1 --port 8000from wavemind import WaveMind
memory = WaveMind(db_path="./state/wavemind.sqlite3")
memory.remember(
"The user prefers short practical answers.",
namespace="user:42",
tags=["preference"],
)
hits = memory.query("How should I answer this user?", namespace="user:42", top_k=3)
for hit in hits:
print(hit.score, hit.text)The integration pattern is intentionally small:
- Call
query()before your app, agent, tool, or UI needs context. - Pass the returned memories into your prompt, screen, search result, or decision function.
- Call
remember()after something worth keeping happens.
The FastAPI server is included in the base install:
wavemind --db ./state/wavemind.sqlite3 serve --host 127.0.0.1 --port 8000Then use WaveMind from any language:
curl -X POST http://127.0.0.1:8000/remember \
-H "Content-Type: application/json" \
-d "{\"text\":\"Andrey prefers short answers\",\"namespace\":\"user:42\",\"tags\":[\"preference\"]}"
curl -X POST http://127.0.0.1:8000/query \
-H "Content-Type: application/json" \
-d "{\"query\":\"How should I answer?\",\"namespace\":\"user:42\",\"top_k\":3}"
curl -X POST http://127.0.0.1:8000/feedback \
-H "Content-Type: application/json" \
-d "{\"id\":1,\"namespace\":\"user:42\",\"useful\":true,\"strength\":0.5,\"reason\":\"used in answer\"}"
curl -X POST http://127.0.0.1:8000/feedback/batch \
-H "Content-Type: application/json" \
-d "{\"namespace\":\"user:42\",\"items\":[{\"id\":1,\"useful\":true,\"strength\":0.5},{\"id\":2,\"useful\":false,\"strength\":0.25}]}"
curl -X POST http://127.0.0.1:8000/forget/batch \
-H "Content-Type: application/json" \
-d "{\"items\":[{\"text\":\"Andrey prefers short answers\",\"namespace\":\"user:42\"}]}"The same feedback loop is available from the CLI:
wavemind --db ./state/wavemind.sqlite3 feedback --id 1 --namespace user:42 --strength 0.5 --reason "used in answer"
wavemind --db ./state/wavemind.sqlite3 feedback-batch --file feedback.jsonWaveMind is local-first. The SQLite database stores memories, vectors, metadata, namespaces, tags, TTL, hotness, priority, and audit events.
| runtime | Suggested database path |
|---|---|
| quick CLI experiment | ./wavemind.sqlite3 |
| Python app or agent | ./state/wavemind.sqlite3 |
| desktop app | user data directory, for example %APPDATA% or ~/.local/share |
| server daemon | /var/lib/wavemind/wavemind.sqlite3 |
| Docker | mounted volume, for example /data/wavemind.sqlite3 |
Explicit path:
wavemind --db ./state/app_memory.sqlite3 remember "Andrey prefers short answers" --namespace user:42
wavemind --db ./state/app_memory.sqlite3 query "answer style" --namespace user:42| You are building... | Start with... |
|---|---|
| Python app | from wavemind import WaveMind |
| LangChain agent | WaveMindMemory from wavemind.integrations.langchain |
| LangGraph workflow | make_recall_node() and make_persist_node() |
| LlamaIndex pipeline | WaveMindRetriever |
| CrewAI or AutoGen loop | The adapters in wavemind.integrations |
| Node, Go, Ruby, PHP, or no-code app | wavemind serve and the HTTP API |
| Personal knowledge base | Store notes by project namespace and query locally |
| Support or CRM workflow | Customer issues, resolutions, preferences, corrections, TTL, and namespace isolation. See examples/customer_support_memory.py. |
| Research or analyst notebook | Findings, hypotheses, decisions, source metadata, TTL, and project isolation. See examples/research_notebook_memory.py. |
For migrations from existing local vector memory, start with
docs/CHROMA_MIGRATION.md. The guide has a tested
offline fixture at examples/chroma_migration.py.
from wavemind import WaveMind
memory = WaveMind(db_path="./state/agent.sqlite3")
def run_turn(user_id: str, user_text: str) -> str:
namespace = f"user:{user_id}"
hits = memory.query(user_text, namespace=namespace, top_k=5, min_score=0.25)
recalled = "\n".join(f"- {hit.text}" for hit in hits)
answer = call_your_llm(f"Relevant memory:\n{recalled}\n\nUser: {user_text}")
memory.remember(f"User said: {user_text}", namespace=namespace, tags=["conversation"])
memory.remember(f"Assistant answered: {answer}", namespace=namespace, tags=["conversation"])
return answerFrom a cloned repository:
$ python examples/demo.py
[ok] Remembered: "Andrey is a trader who tracks market breakouts."
[ok] Remembered: "Andrey prefers short practical answers about product decisions."
Query: "Andrey trader preferences"
-> Result 1 (0.60): "Andrey is a trader who tracks market breakouts."
-> Result 2 (0.30): "Andrey prefers short practical answers about product decisions."
The demo is offline, keyless, and uses the built-in hash encoder.
To see the behavior that plain vector search does not provide:
python examples/dynamic_memory_demo.pyThat demo shows corrected facts outranking stale facts, temporary memory expiring, namespace isolation, and index-health reporting.
To see the same behavior in a practical support/CRM workflow:
python examples/customer_support_memory.pyThat demo stores customer preferences, billing tickets, stale CRM data, temporary discount codes, and separate customer namespaces.
To see source-aware research memory:
python examples/research_notebook_memory.pyThat demo stores analyst findings, temporary hypotheses, decisions, source metadata, and isolated project namespaces.
flowchart LR
A["Text, event, note, document, or agent turn"] --> S["remember()"]
S --> D[("SQLite: text + metadata + vectors + memory state")]
Q["query()"] --> K["k-NN candidate search"]
D --> K
K --> W["wave-field re-rank"]
W --> R["small ranked recall set"]
R --> P["app, search UI, prompt, API, or tool"]
P --> F["recall feedback updates hotness / priority"]
F --> D
F --> C["consolidate active clusters"]
C --> D
The wave field is the dynamic layer around stored memories. It is not a replacement for embeddings; it is the policy that decides which candidate memories should still matter.
| signal | Plain meaning | Effect |
|---|---|---|
| vector similarity | This text is semantically close to the query. | Gets into the candidate set. |
| hotness | This memory has been useful before. | Moves upward during recall. |
| decay | This memory has not mattered recently. | Slowly loses influence. |
| priority | The app says this fact is important. | Raises ranking even before repetition. |
| TTL | This fact is temporary. | Drops out after expiry. |
| namespace and tags | This belongs to one user/project/type. | Prevents cross-user or cross-topic leakage. |
| graph dynamics | Related memories can excite or inhibit each other. | Helps clusters and corrections behave like memory, not a flat list. |
| consolidation | Active clusters can become durable concept memories. | Turns repeated patterns into inspectable higher-level memories with provenance. |
Technically, the current MemoryFieldGraph is a discrete graph over stored
memories, not a continuous mathematical physics field. That honesty matters:
WaveMind is useful today as a dynamic memory engine, while the research path is
to make the field dynamics more explicit, measurable, and scalable.
Self-organization is now part of the core surface. consolidate_concepts(),
wavemind consolidate, and POST /consolidate can turn an active graph cluster
into a new stored memory such as Consolidated memory: systems... without an
LLM call. The generated memory keeps the source memory ids in metadata, so it is
auditable instead of being a hidden summary.
For sentence-transformer embeddings:
python -m pip install "wavemind[sentence]"
wavemind --encoder sentence remember "Andrey is a trader" --namespace demo
wavemind --encoder sentence query "What does Andrey do?" --namespace demoThe default index is NumPy exact search. It is simple and reliable for local memory. For larger candidate generation, WaveMind also exposes optional index backends:
| index | Install | Notes |
|---|---|---|
numpy |
default | Exact cosine search, local, linear scan. |
quantized |
default | Local int8-compressed candidate index with int32-safe scoring. Useful for memory-footprint experiments; approximate recall and latency must still be measured per workload. |
annoy |
pip install "wavemind[indexes]" |
Local ANN. Faster at larger N, but recall must be checked. |
faiss |
pip install "wavemind[indexes]" |
FAISS flat inner-product path where faiss-cpu is available. |
faiss-persisted |
pip install "wavemind[indexes]" |
FAISS with an explicit persisted index snapshot and id map. |
pgvector |
pip install "wavemind[postgres]" |
PostgreSQL/pgvector candidate index. SQLite can still remain the local source of truth. |
qdrant |
pip install "wavemind[indexes]" |
Qdrant service/local-mode candidate index. SQLite remains the source of truth; Qdrant stores vectors. |
Persisted FAISS setup:
export WAVEMIND_FAISS_PATH="./state/wavemind.faiss"
wavemind --index faiss-persisted remember "Andrey is a trader" --namespace demo
wavemind --index faiss-persisted query "trader" --namespace demoSQLite or Postgres remains the source of truth. The persisted FAISS files are a candidate-index snapshot and are validated against the current memory ids, vector dimension, vector count, and a SHA-256 checksum of normalized source vectors on load. If the snapshot does not match the stored memories, WaveMind rebuilds it from the durable store. You can also check and rebuild the candidate index explicitly:
wavemind --index faiss-persisted index-health --json
wavemind --index faiss-persisted rebuild-indexIndex health compares durable memory ids against the candidate index. Local indexes report exact missing/extra ids; service backends report exact ids when the backend exposes an id scan and otherwise fall back to count-based health.
pgvector setup:
export WAVEMIND_PGVECTOR_DSN="postgresql://user:password@localhost:5432/wavemind"
wavemind --index pgvector remember "Andrey is a trader" --namespace demo
wavemind --index pgvector query "trader" --namespace demoOptional pgvector environment variables:
WAVEMIND_PGVECTOR_TABLE- table name, defaultwavemind_vectors.WAVEMIND_PGVECTOR_COLLECTION- collection key, defaultdefault.WAVEMIND_PGVECTOR_CREATE_HNSW=1- create an HNSW index usingvector_cosine_opswhen the installed pgvector version supports it.WAVEMIND_PGVECTOR_HNSW_M- optional HNSW graph degree for index creation.WAVEMIND_PGVECTOR_HNSW_EF_CONSTRUCTION- optional HNSW build accuracy setting.WAVEMIND_PGVECTOR_EF_SEARCH- optional per-query HNSW search depth. Increase it when pgvector is fast but recall is too low.WAVEMIND_PGVECTOR_ITERATIVE_SCAN=strict_order|relaxed_order|off- optional pgvector iterative HNSW scan mode for higher recall on newer pgvector builds.WAVEMIND_PGVECTOR_MAX_SCAN_TUPLESandWAVEMIND_PGVECTOR_SCAN_MEM_MULTIPLIER- optional HNSW scan bounds for production recall/latency tuning.WAVEMIND_PGVECTOR_EXACT=1- force an exact scan for recall audits and correctness-sensitive jobs. This is slower than HNSW, but it gives a direct way to separate index approximation loss from WaveMind ranking behavior.
If WAVEMIND_PGVECTOR_DSN is missing, WaveMind raises a clear error instead of
silently falling back to another index backend.
The pgvector table is created with the current encoder dimension, so use a
separate table when switching between different vector sizes.
Qdrant setup:
export WAVEMIND_QDRANT_URL="http://localhost:6333"
export WAVEMIND_QDRANT_COLLECTION="wavemind_vectors"
wavemind --index qdrant remember "Andrey is a trader" --namespace demo
wavemind --index qdrant query "trader" --namespace demoFor local experiments you can set WAVEMIND_QDRANT_URL=":memory:", but
production latency and durability should be measured against a real Qdrant
service. If WAVEMIND_QDRANT_URL is missing, WaveMind raises a clear error
instead of silently falling back to another backend.
WaveMind now includes an explicit scale preflight:
wavemind scale-plan --target-memories 50000For JSON output in CI or deployment checks:
wavemind --db ./state/wavemind.sqlite3 scale-plan --target-memories 50000 --jsonTo fail a deployment preflight when the plan needs action:
wavemind --db ./state/wavemind.sqlite3 scale-plan --target-memories 50000 --fail-on action_required --jsonIf you only want a plan for a future size without loading optional index packages:
wavemind --index faiss scale-plan --current-memories 10000 --target-memories 50000 --jsonThe scale plan reports:
| field | meaning |
|---|---|
tier |
small, medium, large-local, production-service, or million-plus. |
status |
ok, watch, action_required, or architecture_required. |
recommended_index |
The candidate-index class to use before growth. |
warnings |
Why the current path may fail at the target size. |
actions |
Concrete setup, benchmark, rebuild, and index-health steps. |
The same scale preflight is available over HTTP:
curl "http://127.0.0.1:8000/scale-plan?target_memories=50000"For large production benchmark runs, generate the run contract before starting heavy ingest:
wavemind production-scale-plan --jsonThe checked-in deterministic artifact is:
wavemind production-scale-plan \
--disk-free-gb 0 \
--runner-storage-root state/production-runs \
--write-artifact \
--output benchmarks/production_scale_run_plan.json \
--jsonIt covers the next large-N profiles without claiming unfinished benchmark results:
| profile | engine | target memories | required env | output artifact |
|---|---|---|---|---|
qdrant-10m |
qdrant-service |
10000000 | WAVEMIND_QDRANT_URL |
benchmarks/production_streaming_load_qdrant_10m_results.json |
qdrant-sharded-10m |
qdrant-sharded-service |
10000000 | WAVEMIND_QDRANT_URLS |
benchmarks/production_streaming_load_qdrant_sharded_10m_results.json |
pgvector-10m |
pgvector-service |
10000000 | WAVEMIND_PGVECTOR_DSN |
benchmarks/production_streaming_load_pgvector_10m_results.json |
faiss-ivfpq-50m |
faiss-ivfpq-persisted |
50000000 | WAVEMIND_FAISS_IVFPQ_PATH |
benchmarks/production_streaming_load_ivfpq_50m_results.json |
qdrant-sharded-100m |
qdrant-sharded-service |
100000000 | WAVEMIND_QDRANT_URLS |
benchmarks/production_streaming_load_qdrant_sharded_100m_results.json |
Each profile includes exact command, checkpoint path, runner storage root,
required environment, local runner storage estimate, application storage
estimate, SLO capacity envelope, and cost envelope. Use
--runner-storage-root /mnt/fast/wavemind-runs or
WAVEMIND_PRODUCTION_RUNNER_ROOT to keep checkpoints, resumable ingest state,
and local FAISS indexes off a small system disk. The cost envelope includes
target monthly budget, budget headroom, monthly cost per 1M memories, compute
cost per 1M queries, and machine-readable cost blockers. Override the profile
budget gates with --monthly-budget-usd, --max-cost-per-1m-memories-usd, and
--max-compute-cost-per-1m-queries-usd when planning a specific cluster. A
profile stays action_required until the service/index backend, runner storage
requirements, and cost gates are satisfied. The artifact is a preflight
contract, not latency or recall evidence.
The same artifact also computes a plan-only Pareto frontier across capacity, recall target, p99 target, monthly cost per 1M memories, and compute cost per 1M queries. This helps choose which large-N run should be prioritized first, while keeping the claim boundary explicit: the frontier is planning guidance until the matching output artifact is produced by a real run.
For a concrete operator checklist, use the architecture advisor:
wavemind advise --target-memories 2000000 --namespace-count 4096 --deployment production --replication-factor 3 --read-quorum 1 --read-fanout 1 --jsonIt combines live stats(), scale-plan, p99 targets, namespace count,
replication settings, and multimodal needs into actionable recommendations:
candidate index, sharding, cache, DR drill, observability, and external cluster
load checks. For latency-sensitive hot paths, keep read_fanout near
read_quorum; wider fanout reads more replicas and should be justified by the
HTTP cluster load benchmark. The same advice is available through HTTP:
curl "http://127.0.0.1:8000/architecture/advice?target_memories=2000000&namespace_count=4096&deployment=production&replication_factor=3&read_quorum=1&read_fanout=1"Use --fail-on action_required in CI when a deployment must not proceed until
the required architecture work is done.
For production load tests, use the same SLO and cost gates that power the checked-in benchmark report:
from wavemind import (
ProductionCostTarget,
ProductionSLOTarget,
estimate_production_cost,
evaluate_production_slo,
)
target = ProductionSLOTarget(target_recall_at_k=0.95, target_p99_ms=100, target_qps=100)
result = evaluate_production_slo(
engine="faiss-persisted",
recall_at_k=1.0,
avg_latency_ms=39.12,
p99_latency_ms=57.71,
target=target,
)
print(result.status, result.blocking_reasons)
cost = estimate_production_cost(
result,
memory_count=1_000_000,
vector_dim=128,
target=ProductionCostTarget(
replica_hourly_cost_usd=0.25,
monthly_budget_usd=1500.0,
max_cost_per_1m_memories_usd=500.0,
max_compute_cost_per_1m_queries_usd=10.0,
),
)
print(
cost.cost_status,
cost.required_replicas,
cost.monthly_total_cost_per_1m_memories_usd,
cost.compute_cost_per_1m_queries_usd,
)Rule of thumb:
| target memories | recommended path |
|---|---|
| up to 1000 | SQLite + NumPy exact index. |
| 1000 to 5000 | NumPy can work, but benchmark real queries. |
| 5000 to 50000 | Persisted FAISS for local single-node, or Qdrant service. |
| 50000 to 1M | Service-backed candidate index, namespace sharding, measured p95/p99. |
| above 1M | External vector database plus WaveMind as the memory-policy layer. |
Scale readiness profile:
python benchmarks/scale_readiness_benchmark.py --simulated-memories 1000000Checked-in result:
| profile | result |
|---|---|
| Cluster planner | 4096 namespaces, 4 nodes, replication factor 2, node-loss availability 1.000, zone-loss availability 1.000, write quorum 2, Kubernetes StatefulSet + repair CronJob covering 4096 namespaces. |
| Cluster autoscaler | 10M target memories, RF=3, current nodes 4, required nodes 50, additional nodes 46, target max node load 678711, headroom pass true, full namespace rebalance plan ready: 4094 moves, 82 rolling batches, write quorum 2, checkpoint/repair/validation required for every batch. |
| Control-plane consensus | Majority leadership lease blocks stale leaders, stale revisions, and minority config commits; membership change advances voters 3 -> 5, term 1 -> 2, final config revision 2. |
| Kubernetes operator | CRD + operator deployment true, reconciled StatefulSet, HorizontalPodAutoscaler, rebalance ConfigMap, repair CronJob, and Memory OS CronJob; 10M capacity target raises StatefulSet/HPA to 34 replicas with target max node load 678711, publishes a full rolling rebalance plan with 4048 moves and 81 batches, CPU+memory metrics, production admission target 10000000, status phase Ready, and resources/capacity/autoscaling/rebalance/repair/Memory OS/production-admission/control-plane conditions true. The operator-rendered Memory OS job calls /memory-os/plan before /memory-os/run, applies planned distributed-lock requirements, and blocks mutation when Redis is required but missing. |
| Kubernetes serverless lifecycle | Three stateless API replicas across three kind zones use PVC-backed PostgreSQL, Qdrant, and Redis. Two scale-to-zero cycles restore all 24/24 memories, cross-replica write visibility is 3/3 in 1130.69 ms, delete suppression is 3/3 in 915.03 ms, and a 120-request burst passes with p99 1461.46 ms and zero errors. |
| Kubernetes PostgreSQL/Qdrant DR | A checksummed pg_dump archive (1,016,635 bytes) is restored after source state services stop into fresh PVCs in an independent namespace. Recall is 24/24, the initially empty Qdrant index rebuilds exactly to 24/24, recall remains 24/24 after recovery API pod replacement, and the restore completes in 20.45 s. |
| Hot cache | 2000 lookups, hit rate 0.920, p99 lookup 0.003 ms, query-audit prewarm warmed 1 hot query, prewarm hit true. |
| Query-vector cache | 200 repeated queries, one local encoder call, local hit rate 0.995, Redis-compatible cache shared across workers true; FastAPI service path reuses the encoded query vector and exposes cache hit/miss metrics. |
| API batch query | FastAPI /query/batch answers 100 recall queries in 1 HTTP request instead of 100, preserves vector-cache reuse with one encoder call and batch hit rate 0.990, and exposes batch/cache metrics. |
| Shared rate limiter | Redis-compatible fixed-window limiter, 2 workers, 4 allowed requests, 1 limited request, shared enforcement true. |
| Redis hot cache | Redis-compatible shared cache is visible across workers, query-audit prewarm warms 1 hot query, Memory OS warms 2 observed hot queries plus 6 predictive queries, learns the observed budget recall -> risk limits transition, applies 8 useful/not-useful recall feedback events, demotes cold memories, emits typed self-improvement suggestions plus a policy manifest with 6 decisions, cross-worker hit true, namespace invalidation true, production architecture advice architecture_required. |
| API cache mutation safety | FastAPI shared cache invalidates on /remember, /feedback, and /forget, preventing stale cached recall after memory mutations and after rejected recall feedback. |
| Batch feedback | FastAPI /feedback/batch accepts multiple recall signals in one request, rejects wrong-namespace items, writes audit events, updates positive/negative priority, and invalidates the affected namespace cache once. |
| Distributed sharding | 3 service nodes, replication factor 2, write quorum 2, writes 2, recall after primary loss true, service-mode repair copied 1 missing record, recall after repair true, replicated forget deletes 2, service-mode tombstone suppression before repair true, tombstone repair deleted 1 stale replica record, suppression after repair true, anti-entropy worker repaired 1 missing record and deleted 1 stale tombstone record, query-after-primary-loss 0.84 ms. |
| Distributed HTTP sharding | 3 real localhost API nodes, proxy bypass true, quorum writes 2, recall after primary loss true, HTTP repair copied 1 missing record, recall after repair true, tombstone repair deleted 1 stale API record, suppression after repair true, concurrent writes 12, concurrent query hit rate 1.000. |
| Sustained HTTP cluster load | 4 real localhost API nodes, RF=3, 8 quorum writes through one distributed batch, 8 normal queries, 8 failover queries, 4 deletes through one distributed batch, write HTTP requests reduced 24 -> 4, forget+tombstone HTTP requests reduced 24 -> 8, query HTTP requests reduced 8 -> 3, failover query HTTP requests reduced 8 -> 2, success rate 1.000, failover hit rate 1.000, delete suppression 1.000, repair copied 1 missing replica, p99 operation 389.98 ms. |
| Local HTTP cluster health | Post-load /stats probe reports health true, healthy nodes 4, degraded nodes 0, unavailable nodes 0. |
| Replicated runtime | 3 physical WaveMind stores, replication factor 3, write quorum 2, node-loss recall true, repair copied 1 missing record, tombstone repair deleted 1 stale record, concurrent writes 12, concurrent query hit rate 1.000, p99 query-after-loss 1.34 ms. |
| Active-active delta sync | 2 regions, bidirectional convergence true, full sync imported 6 records, cursor-based incremental sync exported 1 new record and imported 3 replicas, field-only hotness delta exported 0 records and 1 field key, stale import suppressed after delete true, tombstone convergence true, sync 114.70 ms. |
| Sustained active-active sync | 3 independent regions, 3 namespaces, 18 writes, 5 mesh sync cycles, 90 region-pair syncs, cursor count 18, records imported 108, tombstones imported 6, deleted records 6, field keys exported 348, final no-op imported 0, convergence 1.000, delete suppression 1.000, success 1.000, failed pairs 0. |
| HTTP active-active service-region sync | 3 FastAPI service-boundary regions, 2 namespaces, 6 writes, 4 sync cycles, 48 export/import pair calls through /namespace-delta/export and /namespace-delta/import, cursor count 12, records imported 36, tombstones imported 6, deleted records 6, final no-op imported 0, convergence 1.000, delete suppression 1.000, success 1.000, failed pairs 0. |
| Real HTTP active-active service-region smoke | 3 real localhost API region processes, each serving a replicated runtime, 2 namespaces, 6 writes, 3 sync cycles, 36 export/import pair calls, cursor count 12, records imported 36, tombstones imported 6, deleted records 6, final no-op imported 0, convergence 1.000, delete suppression 1.000, success 1.000, failed pairs 0, p99 operation 347.58 ms, SLO true. |
| Field-state CRDT | 3 regions, commutative convergence true, idempotent re-merge true, tombstone-wins true, top-key convergence true, actor watermark convergence true, watermark actors 3, health pass, missing actor detection true, lag detection true, max watermark 100.0, merge 0.13 ms. |
| Replicated snapshot job | 3 replica files, manifest checksum validation true, offsite mirror validation true, portable archive validation true, S3-compatible upload validation true, latest remote archive metadata validation true, remote archive download validation true, object-store DR drill true, object-store retention pruned 2, archive restore 64.13 ms. |
| Structured payloads | image/audio/video/3D/table/event/graph retrieval through the standard memory API, precision@1 1.000; cross-modal target-modality retrieval over persisted payload vectors, precision@1 1.000, vector persistence 1.000, provenance rate 1.000, embedding dim 64; strict external/precomputed vectors for image/audio/video/3D, precision@1 1.000, vector persistence 1.000; external encoder contract over image/audio/table/event/video/3D/graph payloads passes with target precision@1 1.000, global precision@1 1.000, normalized finite persisted vectors 1.000, provenance 1.000, and separation margin 0.811; temporal event retrieval covers actor filters, interval overlap, around-time reranking, recency reranking, persistence, and provenance with precision@1 1.000; knowledge-graph memory covers entity/predicate filters, 2-hop/3-hop traversal, persistence, and provenance with precision@1 1.000, path precision@1 1.000. |
| 100M capacity envelope | 100000000 target memories, 32768 deterministic namespace buckets, weighted rendezvous zone-aware placement, 128 nodes, 8 zones, replication factor 3, node-loss availability 1.000, zone-loss availability 1.000, distinct replica rate 1.000, zone-spread rate 1.000, replica-load skew 1.094, max storage per node 5.81 GB; scale-out audit from 128 to 160 nodes adds 32 nodes, moves 0.492 of replica sets, keeps target replica skew 1.082, and keeps target zone-spread 1.000; valid capacity plan true. |
This profile validates routing, cluster autoscale planning, full rolling
rebalance planning, control-plane majority lease/config revision safety,
Kubernetes deployment, HPA autoscaling, operator status conditions including
RebalancePlanned, MemoryOSReady, ProductionAdmissionReady, and
ControlPlaneReady, and scheduled repair
manifest generation, service-mode distributed namespace sharding, real HTTP
shard transport, sustained mixed HTTP cluster load, replica
repair and tombstone-aware delete repair, plus a reusable anti-entropy repair
worker, quorum-replicated runtime behavior, query-audit cache prewarm,
query-vector cache, Redis-compatible shared rate limiting,
Redis-compatible shared cache behavior, Memory OS shared prewarm, explicit useful/not-useful recall feedback, batch feedback updates, transition-learned predictive prefetch, typed self-improvement suggestions, machine-readable policy decisions for prefetch/priority/forgetting/consolidation/scale/coordination, production architecture advice, and namespace
invalidation, API cache mutation safety on remember/feedback/feedback-batch/forget, cursor-based active-active namespace
delta sync, sustained active-active mesh sync, HTTP service-region
active-active sync, real multi-process active-active service-region smoke,
field-only hotness delta sync,
field-state CRDT convergence with actor watermarks plus missing/lag diagnostics,
replicated snapshot/restore, structured payload handling,
deterministic cross-modal payload retrieval, external precomputed-vector
compatibility, and provenance,
and a 100M-memory capacity-planning envelope,
including verified offsite, archive, object-store latest lookup, object-store
download, object-store retention, and a disaster recovery drill that restores
the latest object-store archive, disables the restored primary replica, and
confirms recall still works from the remaining replicas.
It is not a 10M-vector load test. Real 100k, 1M, and 10M latency claims should
come from service-backed FAISS/Qdrant/pgvector load tests on production-like
hardware.
Local HTTP cluster smoke:
python benchmarks/local_http_cluster_smoke.py \
--nodes 4 \
--replication-factor 3 \
--read-fanout 1 \
--namespace-count 4 \
--memories-per-namespace 2 \
--workers 4 \
--timeout 3 \
--fail-on-sloThis starts 4 real localhost WaveMind API processes with isolated SQLite files,
runs the same service-mode workload through HTTP, and fails if quorum writes,
queries, simulated node failover, missing-replica repair, replicated forget, or
delete suppression regress. The checked-in run reaches success 1.000,
failover hit 1.000, delete suppression 1.000, repaired replicas 1, and
p99 257.13 ms.
Local HTTP active-active service-region smoke:
python benchmarks/local_http_active_active_smoke.py \
--regions 3 \
--replicas-per-region 3 \
--namespace-count 2 \
--timeout 3 \
--fail-on-sloThis starts 3 real localhost WaveMind API region processes. Each region serves a
replicated local runtime through FastAPI, then the runner exchanges namespace
deltas through /namespace-delta/export and /namespace-delta/import. The
checked-in run reaches convergence 1.000, delete suppression 1.000, pair
sync success 1.000, final no-op imports 0, p99 operation 347.58 ms, and
SLO true.
External URL-based active-active loopback:
python benchmarks/external_http_active_active_loopback.py \
--regions 3 \
--replicas-per-region 3 \
--namespace-count 16 \
--timeout 3 \
--fail-on-sloThis starts 3 real localhost WaveMind API regions, passes their URLs into the
same external active-active runner used for remote deployments, and verifies the
URL-based transport contract. The checked-in run reaches convergence 1.000,
delete suppression 1.000, success 1.000, final no-op imports 0, p99
operation 349.21 ms, and SLO true. It proves the external-runner contract,
not remote Kubernetes/serverless operation.
External HTTP active-active regions:
python benchmarks/local_http_active_active_smoke.py \
--region us-east=https://us-east.example.com \
--region eu-west=https://eu-west.example.com \
--region ap-south=https://ap-south.example.com \
--deployment-id staging-active-active-2026-07-07 \
--environment staging \
--source k8s-service \
--namespace-count 16 \
--fail-on-slo \
--output benchmarks/external_http_active_active_results.jsonThe same profile is available as the manual GitHub Actions workflow
external-http-active-active. It is intentionally tracked as non-gating
external evidence until a real remote artifact is committed; the local service
smoke above is not treated as proof of remote Kubernetes/serverless operation.
External HTTP cluster load:
python benchmarks/http_cluster_load_benchmark.py \
--nodes-file deploy/cluster/external-http-cluster.sample.json \
--replication-factor 3 \
--read-quorum 1 \
--read-fanout 1 \
--namespace-count 32 \
--memories-per-namespace 8 \
--workers 8 \
--batch-query-size 24 \
--fail-on-sloThis runs the same mixed workload against user-supplied API nodes: quorum
writes, normal queries, simulated node failover queries, missing-replica repair,
replicated forget, delete suppression, external /query/batch recall, p99
latency, and an explicit SLO verdict. Online query p99 and bulk lifecycle batch
p99 are reported separately so ingestion cost cannot be mistaken for recall
latency.
Use this before claiming that a deployment is production-ready outside the
local readiness smoke profile.
deploy/cluster/external-http-cluster.sample.json defines the repeatable node
manifest shape with deployment id, environment, source, node URLs, and zones.
Core readiness can pass without this artifact, but the strict production gate
only accepts it with non-loopback node addresses, a full Git commit SHA, a
traceable workflow run, and target-specific admission. Kind evidence additionally
requires a SHA-256-linked physical worker-failure artifact over the same pod DNS
endpoints.
The same external-cluster profile can be started from GitHub Actions via
.github/workflows/external-http-cluster-load.yml. Paste one id=https://host
node per line, comma-separated, or semicolon-separated, or paste the node
manifest JSON into nodes_manifest_json. Optionally set the WAVEMIND_API_KEY
repository secret, and set commit_results=true only when the run should
refresh the public benchmark artifacts in main.
Cluster placement planning:
wavemind cluster-plan \
--namespace-count 4096 \
--node node-a=10.0.0.1:8000 \
--node node-b=10.0.0.2:8000 \
--node node-c=10.0.0.3:8000 \
--replication-factor 2 \
--kubernetes \
--repair-cronjob \
--repair-api-key-secret wavemind-api-key \
--jsonThis uses deterministic rendezvous placement so each namespace has a primary
and replica set. The emitted Kubernetes StatefulSet manifest is a deployment
starting point, and the optional repair CronJob runs scheduled service-mode
anti-entropy repair against the same node and namespace plan. Runtime quorum
replication is available through ReplicatedWaveMind; cluster membership and
operator config changes can be guarded with the deterministic
ControlPlaneConsensus / wavemind control-plane-consensus majority lease
preflight. Remote production services should still wrap that contract in a
durable operator/control-plane store.
Control-plane safety preflight:
wavemind control-plane-consensus --jsonThis deterministic profile verifies the operator-side invariants that prevent split-brain config changes: majority leadership lease, stale-leader rejection, stale revision rejection, minority partition rejection, and monotonic config revisions.
Helm deployment:
helm install wavemind ./deploy/helm/wavemindFor authenticated API nodes, create a Secret and reference it:
kubectl create secret generic wavemind-auth --from-literal=admin-key="$WAVEMIND_ADMIN_KEY"
helm upgrade --install wavemind ./deploy/helm/wavemind \
--set auth.enabled=true \
--set auth.existingSecret=wavemind-authThe chart deploys a StatefulSet, normal and headless Services, optional auth
Secret wiring, a scheduled cluster-repair CronJob, and optional Memory OS
CronJobs that call /memory-os/plan before /memory-os/run. It uses
ghcr.io/caspiang/wavemind by default; set image.repository when deploying
from a private registry. The Memory OS CronJob applies the returned plan before
mutation: planned distributed-lock requirements are ORed into /memory-os/run,
and a Redis-required plan fails early if runtime.redisUrl is missing.
helm upgrade --install wavemind ./deploy/helm/wavemind \
--set memoryOs.enabled=true \
--set runtime.auditQueries=1 \
--set runtime.redisUrl=redis://redis.default.svc.cluster.local:6379/0With memoryOs.strictPlan=true, the Memory OS job fails before mutation when
the plan reports architecture_required.
For strict production admission, route the container through wavemind serve
and require checked evidence before the API opens port 8000:
helm upgrade --install wavemind ./deploy/helm/wavemind \
--set productionAdmission.enabled=true \
--set productionAdmission.targetMemories=100000000 \
--set productionAdmission.engine=qdrant-sharded-serviceThis wires WAVEMIND_REQUIRE_PRODUCTION_ADMISSION=1,
WAVEMIND_PRODUCTION_TARGET_MEMORIES, WAVEMIND_PRODUCTION_ENGINE, and
WAVEMIND_PRODUCTION_ADMISSION_ROOT into the StatefulSet. If the required
strict-evidence artifact is missing or rejected, wavemind serve exits before
binding the HTTP port, so a Kubernetes rollout cannot silently bypass the
production evidence gate.
Operator-style deployment:
wavemind operator-bundle --namespace wavemind-system --json | kubectl apply -f -
kubectl apply -f deploy/operator/wavemindcluster.sample.json
wavemind operator-reconcile --file deploy/operator/wavemindcluster.sample.json --out wavemind-resources.json
wavemind operator-status --file deploy/operator/wavemindcluster.sample.json --ready-replicas 3 --json
kubectl apply -f wavemind-resources.jsondeploy/operator contains the WaveMindCluster custom resource path. The
bundle installs the CRD with a status subresource, RBAC, operator Deployment,
and a sample cluster. The reconciler renders the concrete Service, headless
Service, StatefulSet, HPA, rebalance ConfigMap, repair CronJob, and Memory OS
CronJob resources. The Memory OS job calls /memory-os/plan before
/memory-os/run, applies planned distributed-lock requirements, and exits
before mutation when Redis is required but spec.cache.redisUrl is missing.
wavemind operator-status turns the custom resource plus observed
replicas/memory/node health into Kubernetes-style conditions for resources,
capacity, autoscaling, rolling rebalance planning, repair, Memory OS scheduling,
and control-plane safety. spec.controlPlane.consensus is enabled by default
and requires majority leader lease/config revision safety before the cluster is
reported ready. wavemind operator-loop can run in-cluster to keep resources
applied and patch the WaveMindCluster.status subresource when the Kubernetes
client supports it.
The generated operator Deployment runs two replicas with rolling updates and
cross-node anti-affinity. Runtime reconciliation is protected by a durable
Kubernetes Lease stored through the API server/etcd: only the pod holding the
Lease applies resources or patches status, renewals use resourceVersion CAS,
and an expired holder can be replaced with an audited transition counter.
The kubernetes-operator-smoke workflow exercises that path in a real
four-node kind cluster by deleting the leader and a data pod, then verifying
Lease takeover, post-failover reconcile, StatefulSet scaling, and API recovery.
It is CI evidence, not a substitute for the required remote-cluster artifact.
The checked-in result passed all 14/14 checks, including PDB/topology
protection and a CR-driven rolling upgrade that replaced all four data pods,
and links back to the exact
GitHub Actions run.
The operator exposes the same production admission contract through
spec.productionAdmission. Explicitly enable it with:
wavemind operator-sample \
--production-admission \
--production-admission-target-memories 100000000 \
--production-admission-engine qdrant-sharded-service \
--jsonFor capacity targets at or above 10000000 memories, the reconciler also
auto-injects the admission environment into the rendered StatefulSet and reports
ProductionAdmissionReady in operator status.
The operator also accepts capacity targets. Add this to spec.autoscaling and
the reconciler will raise the StatefulSet replicas and HPA min/max replicas to
fit the target under the requested per-node headroom:
{
"enabled": true,
"targetMemories": 10000000,
"maxMemoriesPerNode": 1000000,
"headroom": 0.7,
"rebalance": {
"batchSize": 50,
"maxNodeMovesPerBatch": 50,
"previewBatches": 3
}
}Rendered resources include memory.wavemind.ai/capacity-* annotations with the
calculated replica count and target max node load. They also include
memory.wavemind.ai/rebalance-* annotations plus a bounded
<cluster>-rebalance-plan ConfigMap with full-plan status, move count, batch
count, quorum, checkpoint/repair/validation requirements, and a preview of early
batches.
The same planner is available over HTTP as POST /cluster-plan.
Cluster autoscale planning:
wavemind cluster-autoscale-plan \
--namespace-count 4096 \
--node node-a=https://wm-a.internal \
--node node-b=https://wm-b.internal \
--node node-c=https://wm-c.internal \
--replication-factor 3 \
--target-memories 10000000 \
--max-memories-per-node 1000000 \
--headroom 0.70 \
--zone zone-a --zone zone-b --zone zone-c \
--max-moves 4096 \
--rebalance-plan \
--rebalance-batch-size 50 \
--rebalance-max-node-moves-per-batch 50 \
--jsonThis calculates the required node count for the target memory volume, adds
future nodes with deterministic names and addresses, checks the target max
per-node memory load against the headroom limit, and can emit a rolling
rebalance plan. The rebalance plan groups namespace moves into bounded batches,
tracks read/write quorum, blocks drain-node target violations, and marks every
batch as requiring a source/target checkpoint, cluster repair, and validation
before the next batch. The HTTP surface is POST /cluster-autoscale-plan.
Serverless deployment:
wavemind serverless-sample --namespace wavemind-system --max-scale 256 --out deploy/serverless/wavemind-serverless.sample.json
wavemind serverless-sample --readiness
wavemind serverless-sample --operational-profile --max-scale 256 --target-concurrency 80
wavemind serverless-sample --operational-profile --max-scale 256 --target-concurrency 80 --observed-telemetry deploy/serverless/observed-telemetry.loopback.json
python benchmarks/serverless_observed_telemetry_benchmark.py --node https://wm-a.example --node https://wm-b.example --api-key "$WAVEMIND_API_KEY" --seed-mode first --external-cold-start-ms 900 --output deploy/serverless/observed-telemetry.remote.jsondeploy/serverless contains a stateless API worker plan with two profiles: a
Knative scale-to-zero Service, and a KEDA Deployment/Service/CPU
ScaledObject profile where the autoscaler targets the generated Deployment
and keeps one warm replica. It is
intentionally stricter than local mode: Postgres is required as the source of
truth, Qdrant is used as the external candidate index, Redis is used for shared
hot-query cache, and API keys are read from Kubernetes Secrets. This path is the
current foundation for Knative scale-to-zero and managed/serverless
deployments; CPU-based KEDA handles scale-out, not zero-to-one activation. It
is not a claim that WaveMind has a hosted control plane yet.
The scale-readiness gate also runs a deterministic serverless operational
profile: 3200 requests/second, 80 ms average request time, 320 ms warm p99,
900 ms modeled cold start, 4 required replicas, 256000 burst RPS capacity,
cold-start budget pass, and estimated monthly compute cost $81.76. The
checked-in observed telemetry is generated by
benchmarks/serverless_observed_telemetry_benchmark.py: it starts a balanced
pool of real local WaveMind HTTP API replicas, seeds and warms the hot-query
cache on each replica, measures pool RPS, per-replica RPS, p95/p99/error
rate/cold start, and multiplies measured per-replica throughput by
max_scale=256 for the Knative/KEDA horizontal capacity estimate. This is
loopback evidence, not a real-cluster performance claim. The same runner also
accepts repeated --node https://... URLs, --api-key, --seed-mode first,
and --external-cold-start-ms so the identical telemetry JSON contract can be
run against real Knative/KEDA, Kubernetes, or managed serverless API nodes
before publishing managed serverless numbers.
Maintenance workers:
wavemind maintenance --namespace user:42 --consolidate-steps 10 --consolidate-concepts --json
wavemind memory-os-plan --namespace user:42 --deployment production --target-memories 2000000 --namespace-count 4096 --cache-mode auto --json
wavemind cluster-admission --deployment production --min-nodes 4 --namespace-count 32 --replication-factor 3 --read-quorum 1 --read-fanout 1 --batch-query-size 24 --allow-plan-only --write-artifacts --json
wavemind active-active-admission --deployment production --min-regions 3 --namespace-count 16 --allow-plan-only --write-artifacts --json
wavemind serverless-admission --deployment production --target-rps 3200 --target-p99-ms 500 --max-scale 256 --allow-plan-only --write-artifacts --json
wavemind multimodal-external-evidence --manifest path/to/external_multimodal_manifest.json --write-artifacts --output benchmarks/multimodal_external_encoder_results.json --markdown-output benchmarks/MULTIMODAL_EXTERNAL_EVIDENCE.md --json
wavemind multimodal-admission --deployment production --allow-plan-only --write-artifacts --json
wavemind memory-os-canary --target-memories 100000 --namespace-count 64 --deployment staging --write-artifacts --json
wavemind memory-os-evolution --cycles 3 --write-artifacts --json
wavemind memory-os-admission --target-memories 10000000 --namespace-count 4096 --deployment production --allow-plan-only --write-artifacts --json
wavemind memory-os-policy-bundle --write-artifacts --json
wavemind memory-os --namespace user:42 --redis-url redis://localhost:6379/0 --lock-required --min-frequency 2 --max-hot-queries 32 --json
wavemind cluster-health --node node-a=https://wm-a.internal --node node-b=https://wm-b.internal --node node-c=https://wm-c.internal --replication-factor 3 --read-quorum 1 --read-fanout 1 --api-key "$WAVEMIND_API_KEY" --fail-on-degraded --json
wavemind cluster-repair --node node-a=https://wm-a.internal --node node-b=https://wm-b.internal --node node-c=https://wm-c.internal --namespace user:42 --replication-factor 3 --write-quorum 2 --read-quorum 1 --read-fanout 1 --api-key "$WAVEMIND_API_KEY" --json
wavemind cluster-plan --namespace-count 4096 --node node-a=https://wm-a.internal --node node-b=https://wm-b.internal --node node-c=https://wm-c.internal --replication-factor 3 --repair-cronjob --repair-api-key-secret wavemind-api-key --json
wavemind replicated-snapshot --root ./state/replicas --node node-a --node node-b --node node-c --out ./backups/replicated --offsite ./offsite/replicated --archive ./archives/replicated --s3 s3://my-bucket/wavemind/prod --keep-last 7 --s3-keep-last 30 --json
wavemind replicated-drill --from s3://my-bucket/wavemind/prod --to ./state/drill-restore --query "short support replies" --expect-text "Tenant A prefers short support replies." --jsonThe first command runs one deterministic memory pass: expired-memory purge,
optional field/concept consolidation, and index-health repair. The memory-os
command is the adaptive worker: it reads query audit events, identifies hot
queries, warms Redis/local cache, generates predictive neighbor queries from
the top recalled memories, learns observed follow-up transitions such as
budget recall -> risk limits, predicts bounded priority boosts from usage
patterns, demotes cold unused memories with bounded adaptive forgetting, purges
expired memories, consolidates active clusters into durable concept memories,
checks index health, and returns operator-facing recommendations plus typed
self-improvement suggestions with ids, severity, actions, and evidence for
Studio/operator dashboards. It also emits a policy manifest that turns runtime
signals into explicit decisions for prefetch, priority learning, adaptive
forgetting, consolidation, scale, and distributed coordination. When given
production targets such as --target-memories, --namespace-count,
--deployment production, and --multimodal, it also embeds the same
architecture-advisor output used by release readiness gates: service-index,
namespace-sharding, production-controls, replication capacity, load-test, and
multimodal-readiness actions. In production, use --lock-required with
--redis-url so CronJob retries or multiple workers cannot run overlapping
consolidation, forgetting, and prewarm cycles for the same namespace. The
cluster-health command probes every
WaveMind API node, exposes healthy/degraded/unavailable circuit state, and can
fail deployment preflight when any node is degraded. The cluster-repair command runs service-mode
anti-entropy repair across WaveMind API nodes:
missing replica records are copied back, and tombstoned stale records are
deleted instead of resurrected. The cluster-plan command emits a Kubernetes CronJob
for that repair loop. The snapshot command creates a verified replicated snapshot,
mirrors it to an offsite path,
writes a portable .tar.gz archive, verifies that archive, can upload it to an
S3-compatible object store, verify newest-archive metadata, run an object-store
disaster-recovery drill, and apply local and object-store retention. Production
deployments can call these commands from cron, systemd, Kubernetes CronJobs,
Celery, RQ, or Temporal.
The memory-os-plan command and GET /memory-os/insights endpoint are the
read-only scheduler preflight for that worker set. They inspect current stats,
audited query traffic, and previous Memory OS policy outcomes without mutating
memory, then emit concrete task cadences, worker counts, Redis/shared-cache
requirements, distributed-lock requirements, and exact commands for
memory-os, cache-prewarm, consolidation, forgetting, maintenance, and
architecture-advice loops. The insights endpoint additionally returns
dashboard-ready typed suggestions with ids, severity, action text, and evidence
for Studio and operator consoles. In production mode the planner automatically
promotes --cache-mode auto to Redis when QPS, namespace count, memory count,
hot query volume, or repeated prefetch policy gaps require cross-worker cache
sharing. Plans also include policy_manifest, policy_history,
policy_escalation_ids, and policy_auto_adjustments, so operators can see
when repeated Memory OS gaps changed cadence, priority, cache mode, or lock
requirements. Production memory-os commands emitted by the planner include
--lock-required whenever the plan requires a distributed single-flight lock.
wavemind cluster-admission is the deployment-facing gate for remote
service-node cluster rollout. It joins the strict external_http_cluster
evidence requirement with cluster-node preflight state and writes
benchmarks/cluster_admission_results.json plus
benchmarks/CLUSTER_ADMISSION.md. --fail-on-blocked stops deploys until real
external HTTP service nodes have passed quorum writes, recall, failover,
repair, delete suppression, batch query, and p99 SLO checks; local loopback
HTTP-cluster profiles remain development evidence only.
wavemind active-active-admission is the deployment-facing gate for remote
multi-region active-active rollout. It joins the strict
external_http_active_active evidence requirement with active-active preflight
state and writes benchmarks/active_active_admission_results.json plus
benchmarks/ACTIVE_ACTIVE_ADMISSION.md. --fail-on-blocked stops deploys
until real external HTTP regions have passed convergence, tombstone,
final-noop, and p99 SLO checks; --allow-plan-only keeps the operator report
useful without admitting production traffic.
wavemind serverless-admission is the matching gate for managed/serverless
rollout. It joins the strict serverless_remote_telemetry evidence requirement
with the remote-node preflight state and writes
benchmarks/serverless_admission_results.json plus
benchmarks/SERVERLESS_ADMISSION.md. --fail-on-blocked stops deploys until
real deployed API nodes have produced remote telemetry for p99 latency,
cold-start budget, error rate, and scale-out capacity.
wavemind multimodal-external-evidence turns a real external multimodal
manifest into benchmarks/multimodal_external_encoder_results.json: assets
must already have external shared-space vectors, s3:// object-store metadata,
verified sha256/byte-size provenance, and precomputed query vectors.
wavemind multimodal-admission is the deployment-facing gate for production
multimodal claims. It uses the checked structured-memory report as the API
contract, but only admits production when the external evidence artifact proves
real image/audio/video/3D encoder quality, object-store-backed assets, object
verification, vector persistence, provenance, and query/encode latency SLOs.
Deterministic structured fixtures stay useful for development, but they do not
unlock broad multimodal model-quality claims.
wavemind memory-os-admission is the stricter deployment gate for the same
worker set: it checks hot-query audit signal, Redis/shared-cache wiring,
distributed lock wiring, singleton/idempotent mutations, policy coverage, and
strict architecture boundaries before Memory OS workers become production
automation. It writes benchmarks/memory_os_admission_results.json and
benchmarks/MEMORY_OS_ADMISSION.md; --fail-on-blocked makes CI/deploys stop
until the worker set is really admitted.
wavemind memory-os-canary is the staging proof for that gate: it seeds
representative memories and query-audit traffic, runs one Memory OS cycle,
checks prewarm, predictive prefetch, priority learning, TTL cleanup, and then
verifies that scheduler admission passes when Redis/cache and lock wiring are
declared. It writes benchmarks/memory_os_canary_results.json and
benchmarks/MEMORY_OS_CANARY.md. This is a staging canary, not remote
Kubernetes, real Redis, or 10M/100M production evidence.
wavemind memory-os-evolution is the multi-cycle proof for that same worker
loop. It replays representative query-audit traffic across several Memory OS
cycles, verifies that repeated required policy gaps become history-backed
suggestions and scheduler escalations, and checks that stable OK policies,
hot-query prewarm, predictive prefetch, priority learning, and required worker
tasks remain active. It writes
benchmarks/memory_os_policy_evolution_results.json and
benchmarks/MEMORY_OS_POLICY_EVOLUTION.md. This is deterministic local/staging
policy evidence; it does not unlock unattended production automation without
remote Redis, distributed lock, runtime environment, and strict large-scale
evidence.
wavemind memory-os-policy-bundle connects those artifacts to runtime. It
reads the checked canary, policy-evolution, and admission reports, then emits
benchmarks/memory_os_policy_bundle_results.json and
benchmarks/MEMORY_OS_POLICY_BUNDLE.md: a deterministic operator policy
manifest with enabled Memory OS tasks, required Redis/lock environment,
observability metrics, Kubernetes/CronJob patch data, and explicit promotion
gates. Current checked-in evidence is staging-promotable but keeps production
locked while memory-os-admission is plan_only.
Hot-cache options:
| cache | use case |
|---|---|
HotMemoryCache |
in-process local API/server cache. |
RedisHotMemoryCache |
shared cache for multiple API workers. Install with pip install "wavemind[redis]". |
QueryVectorCache |
in-process cache for encoded query vectors when the encoder is expensive. |
RedisQueryVectorCache |
shared encoded-query-vector cache across API workers. |
API cache can be enabled with:
WAVEMIND_CACHE_CAPACITY=512 WAVEMIND_CACHE_TTL_SECONDS=60 wavemind serveFor multiple API workers, use a shared Redis cache:
WAVEMIND_REDIS_URL=redis://localhost:6379/0 WAVEMIND_AUDIT_QUERIES=1 wavemind serveFor repeated natural-language queries with a semantic encoder, cache encoded query vectors separately from full query results:
WAVEMIND_VECTOR_CACHE_CAPACITY=1024 WAVEMIND_VECTOR_CACHE_TTL_SECONDS=300 wavemind serve
WAVEMIND_VECTOR_CACHE_REDIS_URL=redis://localhost:6379/0 wavemind serveTo verify the live multi-process cache path, including Redis-backed batch query recall, shared query-vector cache hits, and batch recall feedback invalidation, against a real Redis service:
python benchmarks/redis_api_load_benchmark.py --redis-url redis://localhost:6379/0 --workers 2 --requests 40 --batch-size 12 --fail-on-sloFor production workers, enable query audit and prewarm the cache from repeated real queries:
WAVEMIND_AUDIT_QUERIES=1 WAVEMIND_CACHE_CAPACITY=512 wavemind serve
wavemind --audit-queries query "budget preference" --namespace demo
curl -X POST http://127.0.0.1:8000/cache/prewarm -H "x-api-key: $WAVEMIND_ADMIN_KEY" -d '{"min_frequency":2,"max_queries":32}'
wavemind cache-prewarm --redis-url redis://localhost:6379/0 --min-frequency 2 --max-queries 32The same path is available in Python through CachePrewarmWorker. The CLI can
also run with a process-local cache for diagnostics, but production prewarm
should use Redis so warmed entries survive the worker process. Query audit
stores query text, so keep it opt-in for deployments with stricter privacy
requirements.
WaveMind can store non-text memories as structured text plus metadata. This is useful for product events, tables, call transcripts, images, videos, 3D assets, and knowledge graphs while keeping the same query API.
from wavemind import WaveMind, image_payload, remember_payload
memory = WaveMind()
remember_payload(
memory,
image_payload("s3://demo/chart.png", caption="enterprise revenue expansion chart"),
namespace="research",
)
print(memory.query("enterprise expansion chart", namespace="research")[0].metadata)For agent workflows that need one memory layer across payload types, use
CrossModalMemoryLayer. It stores typed payloads through WaveMind, keeps
provenance in metadata, and re-ranks by a shared deterministic descriptor
embedding so queries can target one modality without changing the underlying
storage.
from wavemind import CrossModalMemoryLayer, WaveMind, audio_payload, image_payload
memory = WaveMind()
layer = CrossModalMemoryLayer(memory)
layer.remember(image_payload("s3://demo/chart.png", caption="Q2 revenue chart"), namespace="research")
layer.remember(audio_payload("s3://demo/call.wav", transcript="Customer asked about Q2 revenue"), namespace="research")
for result in layer.query("revenue chart", namespace="research", target_modality="image"):
print(result.modality, result.score, result.provenance)If your application already computes CLIP/audio/video/3D embeddings, use the strict precomputed-vector path. WaveMind stores the vector with the payload and requires a query vector at search time, so there is no hidden descriptor fallback:
from wavemind import CrossModalMemoryLayer, PrecomputedCrossModalEncoder, WaveMind, image_payload
memory = WaveMind()
layer = CrossModalMemoryLayer(
memory,
cross_modal_encoder=PrecomputedCrossModalEncoder(vector_dim=512, name="clip"),
)
layer.remember(
image_payload(
"s3://demo/chart.png",
caption="Q2 revenue chart",
metadata={"cross_modal_vector": image_clip_vector},
),
namespace="research",
)
results = layer.query(
"revenue chart",
namespace="research",
target_modality="image",
query_vector=text_clip_vector,
)Before trusting a production multimodal encoder, run the external-vector contract. It writes representative image, audio, table, event, video, 3D, and graph payloads through the real memory layer, then checks global retrieval, target-modality routing, persisted finite normalized vectors, provenance, and separation margin:
from wavemind import WaveMind, validate_precomputed_cross_modal_contract
memory = WaveMind()
report = validate_precomputed_cross_modal_contract(memory)
assert report.ok, report.failuresThe same contract is part of the scale-readiness and production-readiness gates, so external CLIP/audio/video/3D integrations must prove the storage and recall contract before they become published evidence.
For encoders that produce both payload and query vectors, run the active encoder health check as a deployment preflight. It probes all supported modalities, checks finite normalized vectors, target routing, global precision@1, dimension compatibility, separation margin, and p95 encode latency:
from wavemind import (
DescriptorCrossModalEncoder,
HashingTextEncoder,
check_cross_modal_encoder_health,
)
encoder = DescriptorCrossModalEncoder(HashingTextEncoder(vector_dim=64), vector_dim=64)
report = check_cross_modal_encoder_health(encoder)
assert report.ok, report.failuresThe checked-in structured-memory report now includes this health gate, so a multimodal backend can fail before it reaches production traffic.
For production media, keep large files in S3-compatible object storage and store a verified content-addressed manifest with the memory. This keeps SQLite/Postgres as metadata source of truth while video, audio, image, and 3D bytes live in S3, R2, or MinIO:
from wavemind import S3AssetStore, video_payload
assets = S3AssetStore.from_uri("s3://wavemind-assets/media")
asset = assets.upload_asset("demo.mp4", kind="video")
payload = video_payload(
asset.uri,
summary="memory graph demo",
metadata=asset.payload_metadata(),
)asset.payload_metadata() includes asset_sha256, asset_bytes,
asset_media_type, and asset_verified; cross-modal results return those fields
in provenance so downstream agents can audit the exact media object behind a
recall.
For a built-in CLIP-style image/text backend, install the multimodal extra:
python -m pip install "wavemind[multimodal]"from wavemind import CrossModalMemoryLayer, SentenceTransformersCrossModalEncoder, WaveMind, image_payload
memory = WaveMind()
layer = CrossModalMemoryLayer(
memory,
cross_modal_encoder=SentenceTransformersCrossModalEncoder("clip-ViT-B-32"),
)
layer.remember(
image_payload("chart.png", caption="Q2 revenue chart"),
namespace="research",
)
results = layer.query("revenue chart", namespace="research", target_modality="image")This backend loads local image files with Pillow and encodes text queries through the same sentence-transformers model. Remote assets, audio, video, and 3D payloads should either carry strong text descriptors or use the precomputed vector path until dedicated perception backends are benchmarked.
Supported payload helpers:
| helper | use case |
|---|---|
image_payload() |
image URI plus caption or alt text |
audio_payload() |
audio URI plus transcript or summary |
video_payload() |
video URI plus transcript, scenes, duration, and summary |
asset3d_payload() |
3D model URI plus labels, dimensions, and format |
table_payload() |
compact table preview with row count |
event_payload() |
structured product, user, or system event |
graph_payload() |
knowledge graph triples stored as queryable memory |
For graph-heavy memory, use KnowledgeGraphMemoryLayer when you need entity
filters and multi-hop traversal instead of plain text recall:
from wavemind import KnowledgeGraphMemoryLayer, WaveMind
memory = WaveMind()
graph = KnowledgeGraphMemoryLayer(memory)
graph.remember_triples(
[
("Andrey", "works_on", "trading agent"),
("trading agent", "uses", "WaveMind memory"),
],
namespace="agent",
title="agent memory graph",
)
path = graph.query(
"how is Andrey connected to WaveMind memory?",
namespace="agent",
subject="Andrey",
object="WaveMind memory",
max_depth=2,
)[0]
print(path.depth, path.path, path.provenance)Temporal events can also be queried as time-aware memory:
from wavemind import TemporalEventMemoryLayer, WaveMind
memory = WaveMind()
events = TemporalEventMemoryLayer(memory)
events.remember(
"risk limits reviewed",
namespace="agent:trading",
actor="agent",
timestamp="2026-07-07T12:00:00Z",
tags=["risk"],
)
fresh = events.query(
"risk limits",
namespace="agent:trading",
actor="agent",
recency_anchor="2026-07-08T12:00:00Z",
)SQLite is the default source of truth. For multi-tenant production deployments, WaveMind also exposes PostgreSQL as an explicit source-of-truth backend:
export WAVEMIND_STORE="postgres"
export WAVEMIND_POSTGRES_DSN="postgresql://user:password@localhost:5432/wavemind"
wavemind --store postgres remember "Andrey is a trader" --namespace user:andrey
wavemind --store postgres query "trader" --namespace user:andreyOptional table environment variables:
WAVEMIND_POSTGRES_MEMORIES_TABLE, defaultwavemind_memories.WAVEMIND_POSTGRES_AUDIT_TABLE, defaultwavemind_audit_events.
Postgres storage is separate from pgvector: Postgres storage keeps memories,
metadata, TTL, audit events, and vectors as durable application state; pgvector
is a candidate index backend for nearest-neighbor search. You can use SQLite
storage with pgvector, Postgres storage with NumPy/FAISS/Qdrant, or eventually
Postgres storage plus pgvector when you want both state and vector search inside
PostgreSQL.
Exact one-file backup:
wavemind --db ./state/wavemind.sqlite3 backup --out ./backups/wavemind.sqlite3Timestamped backups with retention:
wavemind --db ./state/wavemind.sqlite3 backup --out ./backups --prefix wavemind --keep-last 7Restore into a new or replacement SQLite file:
wavemind restore --from ./backups/wavemind-20260630-120000.sqlite3 --to ./state/wavemind.sqlite3 --overwriteThe backup command uses SQLite's backup API, so it is safe to run while the
process is alive. Restore is intentionally an explicit command and refuses to
overwrite an existing database unless --overwrite is passed.
SQLite point-in-time recovery:
import time
from wavemind import SQLiteMemoryStore, WaveMind
memory = WaveMind(
db_path="./state/wavemind.sqlite3",
recovery_journal_path="./state/wavemind.recovery.jsonl",
)
memory.remember("Tenant A prefers short support replies.", namespace="tenant:a")
checkpoint = time.time()
memory.remember("Tenant A switched to detailed reports.", namespace="tenant:a")
memory.forget(text="Tenant A prefers short support replies.", namespace="tenant:a")
SQLiteMemoryStore.restore_recovery_journal(
"./state/wavemind.recovery.jsonl",
"./state/restored-at-checkpoint.sqlite3",
until=checkpoint,
)Equivalent CLI:
wavemind --db ./state/wavemind.sqlite3 --recovery-journal ./state/wavemind.recovery.jsonl remember "Tenant A prefers short support replies." --namespace tenant:a
wavemind recovery-restore --from ./state/wavemind.recovery.jsonl --to ./state/restored.sqlite3 --overwrite --jsonFor API/server deployments, set WAVEMIND_RECOVERY_JOURNAL=/data/wavemind.recovery.jsonl.
The recovery journal is append-only JSONL. It records remember, forget, and
purge_expired mutations with the persisted memory id, text, metadata, tags,
vector, and field pattern, so restore replay does not need to call the encoder
again. Use regular SQLite backups for coarse snapshots and the journal when you
need to restore to a specific mutation boundary.
For Postgres storage, use database-native backup tooling such as pg_dump,
managed snapshots, or Postgres point-in-time recovery. WaveMind's JSONL recovery
journal is a local SQLite source-of-truth mechanism, not a replacement for
database-native WAL/PITR in managed Postgres.
Postgres PITR runbook/preflight:
wavemind postgres-pitr-plan --out ./ops/postgres-pitr-plan.json --jsonThis emits a secret-safe database-native runbook with WAL archiving, streaming
pg_basebackup, restore target configuration, replay verification, and
promotion steps. It stores environment variable names only, not DSNs or secret
values. The checked-in artifact is
benchmarks/postgres_pitr_plan.json; a real managed Postgres restore drill
should execute the plan in staging and record replay LSN, target timestamp,
restore duration, and post-restore row/index checks.
Replicated runtime snapshot/restore:
from wavemind import HashingTextEncoder, ReplicatedSnapshotWorker, ReplicatedWaveMind
memory = ReplicatedWaveMind(
root_path="./state/replicas",
nodes=["node-a", "node-b", "node-c"],
replication_factor=3,
encoder=HashingTextEncoder(vector_dim=64),
)
memory.remember("Tenant A prefers short support replies.", namespace="tenant:a")
snapshot_job = ReplicatedSnapshotWorker(memory).run_once(
destination="./backups/replicated",
offsite_destination="./offsite/replicated",
archive_destination="./archives/replicated",
object_store_destination="s3://my-bucket/wavemind/prod",
keep_last=7,
object_store_keep_last=30,
)
assert snapshot_job.ok
restored, report = ReplicatedWaveMind.restore_snapshot_archive(
snapshot_job.archive_path,
"./state/restored-replicas",
encoder=HashingTextEncoder(vector_dim=64),
)The replicated snapshot job writes one SQLite backup per replica plus
manifest.json with SHA-256 checksums, replica metadata, quorum settings, and
node definitions. It can mirror the snapshot to a second path for offsite
backup, write a portable .tar.gz archive for object-store/offsite systems,
verify that archive, upload it to any S3-compatible object store through
boto3, list the newest remote archive, restore from the newest remote archive,
run a disaster-recovery drill from the newest or exact remote archive, and apply
keep_last retention locally, offsite, for archives, and explicitly for
object-store archives through object_store_keep_last.
Restore refuses to overwrite a non-empty root unless overwrite=True is passed.
Equivalent CLI:
wavemind replicated-snapshot \
--root ./state/replicas \
--node node-a --node node-b --node node-c \
--out ./backups/replicated \
--offsite ./offsite/replicated \
--archive ./archives/replicated \
--s3 s3://my-bucket/wavemind/prod \
--keep-last 7 \
--s3-keep-last 30 \
--json
wavemind replicated-s3-archives \
--s3 s3://my-bucket/wavemind/prod \
--latest \
--json
wavemind replicated-restore \
--from ./archives/replicated/wavemind-replicated-20260705-120000.tar.gz \
--to ./state/restored-replicas \
--overwrite \
--json
wavemind replicated-restore \
--from s3://my-bucket/wavemind/prod \
--latest \
--to ./state/restored-replicas \
--overwrite \
--json
wavemind replicated-drill \
--from s3://my-bucket/wavemind/prod \
--to ./state/drill-restore \
--query "short support replies" \
--expect-text "Tenant A prefers short support replies." \
--jsonInstall S3/R2/MinIO support with pip install "wavemind[s3]". For
S3-compatible endpoints such as Cloudflare R2 or MinIO, pass
--s3-endpoint-url and optionally --s3-region.
Run the local FastAPI server:
wavemind --db ./app_memory.sqlite3 serve --host 127.0.0.1 --port 8000Store and query memory over HTTP:
curl -X POST http://127.0.0.1:8000/remember -H "Content-Type: application/json" -d "{\"text\":\"Andrey is a trader\",\"namespace\":\"demo\"}"
curl -X POST http://127.0.0.1:8000/query -H "Content-Type: application/json" -d "{\"query\":\"trader\",\"namespace\":\"demo\",\"top_k\":1}"
curl -X POST http://127.0.0.1:8000/query/batch -H "Content-Type: application/json" -d "{\"queries\":[{\"query\":\"trader\",\"namespace\":\"demo\",\"top_k\":1},{\"query\":\"Andrey\",\"namespace\":\"demo\",\"top_k\":1}]}"Operational endpoints:
curl http://127.0.0.1:8000/stats?namespace=demo
curl http://127.0.0.1:8000/audit?namespace=demo
curl http://127.0.0.1:8000/metrics
curl http://127.0.0.1:8000/observability
curl http://127.0.0.1:8000/index/health
curl "http://127.0.0.1:8000/scale-plan?target_memories=50000"
curl -X POST http://127.0.0.1:8000/index/rebuild
curl -X POST http://127.0.0.1:8000/consolidate -H "Content-Type: application/json" -d '{"namespace":"demo","seed_text":"Rust compiler systems","min_energy":0.01}'
curl -X POST http://127.0.0.1:8000/backup -H "Content-Type: application/json" -d '{"path":"./backups","keep_last":7}'/audit returns mutation events such as remember, forget, backup, and
consolidate_concept. Query audit is opt-in with WAVEMIND_AUDIT_QUERIES=1 because
writing an audit row for every query changes latency. /metrics returns a
Prometheus-compatible text payload without adding a required dependency.
/index/health reports source-of-truth versus candidate-index consistency.
/index/rebuild rebuilds the candidate index from stored active memories and
logs an index_rebuild audit event.
Full observability guide and local Prometheus/OTEL examples:
docs/OBSERVABILITY.md.
OpenTelemetry traces are optional and off by default:
pip install "wavemind[otel]"
export WAVEMIND_OTEL_ENABLED=1
export WAVEMIND_OTEL_SERVICE_NAME=wavemind-api
export WAVEMIND_OTEL_EXPORTER=otlp
export WAVEMIND_OTEL_ENDPOINT="http://localhost:4318/v1/traces"
wavemind --db ./app_memory.sqlite3 serve --host 127.0.0.1 --port 8000Use WAVEMIND_OTEL_EXPORTER=console for local trace inspection. FastAPI
requests are instrumented, and core memory phases such as encode, index search,
graph propagation, reranking, load, and backup create spans when OpenTelemetry
is enabled.
Production API controls are opt-in:
export WAVEMIND_READ_KEYS="read-key"
export WAVEMIND_WRITE_KEYS="write-key"
export WAVEMIND_ADMIN_KEYS="admin-key"
export WAVEMIND_RATE_LIMIT_PER_MINUTE=120
export WAVEMIND_API_SERIALIZE_OPERATIONS=1For multiple API workers, use a shared Redis rate-limit bucket:
export WAVEMIND_RATE_LIMIT_PER_MINUTE=120
export WAVEMIND_RATE_LIMIT_REDIS_URL=redis://localhost:6379/0
export WAVEMIND_RATE_LIMIT_REDIS_PREFIX=wavemind:rateWAVEMIND_API_SERIALIZE_OPERATIONS=1 is the default. It keeps one in-process
FastAPI worker from running concurrent operations through the same local
WaveMind/SQLite runtime. Set it to 0 only when the backing store and index
path are safe for concurrent in-process access.
Role behavior:
| role | Env var | Allows |
|---|---|---|
| read | WAVEMIND_READ_KEYS |
/query, /query/batch, /stats, /metrics, /index/health |
| write | WAVEMIND_WRITE_KEYS |
read actions plus /remember and /import |
| admin | WAVEMIND_ADMIN_KEYS or WAVEMIND_API_KEYS |
all actions, including /audit, /backup, /index/rebuild, /forget, and /forget/batch |
Keys are accepted through Authorization: Bearer <key> or X-API-Key: <key>.
If no key env vars are set, authentication is disabled for local development.
For contributors installing from a local clone:
git clone https://github.com/CaspianG/wavemind.git
cd wavemind
python -m pip install -e ".[sentence]"One-file setup scripts are also included in the repository:
sh install.shinstall.batInstall the optional integration:
pip install "wavemind[langchain]"Use WaveMind as a drop-in LangChain memory object:
from wavemind.integrations.langchain import WaveMindMemory
memory = WaveMindMemory(db_path="agent_memory.sqlite3")
# Replace: memory = ConversationBufferMemory()Offline runnable example from a cloned repository:
python examples/langchain_memory.pyWaveMind only needs two touch points in an agent, service, notebook, or app:
- Before work happens,
query()for relevant memory and pass the short result into the next step: a prompt, search screen, tool call, support workflow, or decision function. - After work happens,
remember()durable facts, preferences, summaries, outcomes, corrections, or notes.
That makes it usable in more than LangChain:
| Use case | Integration style |
|---|---|
| LangChain agent | Use WaveMindMemory from wavemind.integrations.langchain. |
| LangGraph workflow | Use make_recall_node() and make_persist_node() from wavemind.integrations.langgraph. |
| LlamaIndex pipeline | Use WaveMindRetriever from wavemind.integrations.llamaindex. |
| CrewAI crew | Use WaveMindCrewAITools from wavemind.integrations.crewai. |
| AutoGen loop | Use WaveMindAutoGenMemory from wavemind.integrations.autogen. |
| Custom Python agent | Create one WaveMind instance and call query() before the LLM. |
| Node, Go, Ruby, PHP, or no-code app | Run wavemind serve and call the HTTP API. |
| Multi-user SaaS | Use namespace="user:<id>" or namespace="tenant:<id>:agent:<id>". |
| Knowledge base or notebook | Store notes by project namespace and retrieve a small evidence set. |
| Support or CRM workflow | Store issues, preferences, resolutions, and corrections with tags. |
| Research workflow | Store observations with source metadata and expire temporary hypotheses. |
| Temporary context | Store with ttl_seconds=... so stale memory expires automatically. |
| Preference/profile memory | Store with tags such as profile, preference, project, decision. |
| Corrections/privacy | Use forget() or namespace deletion workflows. |
More examples: docs/USE_CASES.md.
Migrating from a Chroma memory store: docs/CHROMA_MIGRATION.md.
Framework examples in this repository:
| Framework / pattern | Example |
|---|---|
| LangChain memory | examples/langchain_memory.py |
| OpenAI/OpenRouter-style agent loop | examples/agent_with_memory.py |
| LangGraph hooks | wavemind.integrations.langgraph, examples/framework_integrations.py |
| LlamaIndex-style retriever | wavemind.integrations.llamaindex, examples/llamaindex_retriever.py |
| CrewAI-style tools | wavemind.integrations.crewai, examples/framework_integrations.py |
| AutoGen-style hooks | wavemind.integrations.autogen, examples/framework_integrations.py |
| Namespace sharding | examples/sharded_memory.py |
Run the dedicated offline LlamaIndex-style retriever example:
python examples/llamaindex_retriever.pyOpenClaw memory is file-centered:
it writes durable memory into MEMORY.md, daily notes under memory/, and uses
tools such as memory_search / memory_get. OpenClaw's documented agent loop
also exposes hooks such as before_prompt_build, agent_end,
message_received, and message_sent.
The safest WaveMind integration is a sidecar, not a replacement:
- Keep OpenClaw's Markdown memory as the human-readable source of durable truth.
- Use WaveMind as the dynamic recall layer for hotness, TTL, namespaces, and correction-sensitive ranking.
- Store the SQLite file outside committed workspace files, for example
~/.openclaw/wavemind/<agent-id>.sqlite3. - Query WaveMind from
before_prompt_buildand inject a compact memory block withprependContext. - Capture new durable summaries from
agent_endor message hooks.
Sketch of the adapter logic:
from pathlib import Path
from wavemind import WaveMind
db_path = Path.home() / ".openclaw" / "wavemind" / "main.sqlite3"
memory = WaveMind(db_path=db_path)
def before_prompt_build(agent_id: str, user_text: str) -> str:
namespace = f"openclaw:{agent_id}"
hits = memory.query(user_text, namespace=namespace, top_k=5, min_score=0.25)
return "\n".join(f"- {hit.text}" for hit in hits)
def agent_end(agent_id: str, summary: str) -> None:
namespace = f"openclaw:{agent_id}"
memory.remember(summary, namespace=namespace, tags=["summary"], priority=1.5)For a production OpenClaw plugin, translate that sketch into the documented
plugin hook surface: before_prompt_build for recall and agent_end /
message_received / message_sent for capture.
The public HERMES Agent is a
LangChain / LangGraph mathematical-reasoning agent. Its README describes
HermesReasoner as a LangChain BaseTool and mentions an optional in-memory
embedding store for previously verified claims.
WaveMind fits there as a persistent memory layer around that loop:
- Recall previously verified claims before
HermesReasoneris invoked. - Store successfully verified claims with
tags=["verified-claim"]. - Scope by
user_id, project, benchmark, or theorem namespace. - Replace short-lived in-memory vector recall when the agent needs restarts, TTL, explicit forgetting, or cross-session reuse.
Generic Hermes-style loop:
from wavemind import WaveMind
memory = WaveMind(db_path="./state/hermes_claims.sqlite3")
def verify_with_memory(user_id: str, problem: str) -> str:
namespace = f"hermes:{user_id}"
claims = memory.query(problem, namespace=namespace, tags=["verified-claim"], top_k=5)
context = "\n".join(f"- {claim.text}" for claim in claims)
result = call_hermes_reasoner(problem=problem, extra_context=context)
if result.label == "CORRECT":
memory.remember(result.claim, namespace=namespace, tags=["verified-claim"], priority=2.0)
return result.textFor any other agent framework, the rule is the same: recall before the model, capture after the turn, isolate users with namespaces, and use TTL for temporary facts.
WaveMind can store any small-to-medium memory stream where meaning, freshness, and repeated use matter. It is useful when "show me the nearest text" is not enough and the application needs "show me what is relevant now."
| Use case | Example |
|---|---|
| Support memory | Recall past user issues, plans, bugs, and resolutions. |
| Product research | Store interview snippets with tags=["customer", "pain"]. |
| Team knowledge | Remember project decisions and suppress expired decisions with TTL. |
| Personal assistant | Store preferences, routines, people, and recurring context. |
| Game/NPC memory | Give characters scoped memory that strengthens after repeated events. |
| Trading research | Store labeled OHLCV pattern notes before building a backtest layer. |
| Document notebook | Import text/PDF/JSON chunks and query by namespace/project. |
| Personal knowledge base | Keep decisions, recurring context, people, links, and notes searchable without sending them to a hosted vector DB. |
WaveMind is not positioned as "a faster Chroma." Chroma, Qdrant, Pinecone, and Weaviate are vector databases: they store embeddings and return nearest neighbors. That is the right tool for many static RAG workloads.
WaveMind is a dynamic memory layer. It still uses vector search first, but then applies memory-specific signals that a plain vector store does not model by default:
| memory behavior | Why it matters | WaveMind mechanism |
|---|---|---|
| Hot memories | Information that keeps being useful should become easier to recall again. | Wave-field hotness and priority updates. |
| Aging memories | Old low-value facts should fade instead of competing forever. | TTL and decay-aware scoring. |
| Scoped memory | One user, app, workspace, or project should not leak into another. | Namespaces and tags. |
| Explicit forgetting | Real systems need deletion, privacy cleanup, and correction workflows. | forget() plus SQLite persistence. |
| Stable restart behavior | A memory system must survive process restarts. | SQLite source of truth, reloadable indexes. |
| Vector plus memory rank | Semantic similarity is necessary but not sufficient for long-running memory. | k-NN candidates first, wave field as re-ranker. |
The current Chroma benchmark below is intentionally conservative: it compares static retrieval on the same facts and the same hash embeddings. That benchmark is useful, but it does not exercise WaveMind's main thesis: memory that changes over time as software recalls, reinforces, ages, and forgets information.
The benchmark that should decide whether WaveMind is worth using is a dynamic memory benchmark:
| scenario | What should happen |
|---|---|
| A fact, preference, or decision is used many times. | WaveMind should rank it higher than equally similar but unused facts. |
| A fact expires via TTL. | WaveMind should suppress it without requiring manual vector cleanup. |
| A user or system corrects an old fact. | WaveMind should prefer the newer or reinforced memory. |
| A query is ambiguous across namespaces. | WaveMind should return only the scoped user's memory. |
| A long history has many irrelevant facts. | WaveMind should preserve useful recall instead of treating all vectors equally. |
In short: static vector search answers "what is nearest?" Dynamic memory also asks "what is still relevant, reinforced, scoped, and allowed to be remembered?"
The main branch stays focused on the core WaveMind library: dynamic memory, storage, indexes, APIs, integrations, and public memory benchmarks.
Experimental domains live in separate branches so they can move quickly without overloading the main README:
| Branch | Scope |
|---|---|
research/crypto-pattern-memory |
OHLCV pattern-memory research, historical analogue retrieval, and future backtest experiments. |
WaveMind tracks benchmarks in two layers:
- Implemented local checks - fast, reproducible scripts that run from this repository and protect the core memory behavior.
- Public benchmark roadmap - external retrieval and memory benchmarks that should decide whether WaveMind is competitive outside hand-made demos.
Machine-readable benchmark matrix: benchmarks/benchmark_matrix_results.json.
Full generated benchmark report: benchmarks/BENCHMARK_REPORT.md.
Compact benchmark leaderboard: benchmarks/BENCHMARK_LEADERBOARD.md.
Agent-impact leaderboard: benchmarks/AGENT_IMPACT.md.
Structured memory report: benchmarks/STRUCTURED_MEMORY.md.
Memory OS intelligence report: benchmarks/MEMORY_OS_INTELLIGENCE.md.
Memory OS policy evolution report: benchmarks/MEMORY_OS_POLICY_EVOLUTION.md.
Memory OS policy bundle: benchmarks/MEMORY_OS_POLICY_BUNDLE.md.
Cluster autoscale report: benchmarks/CLUSTER_AUTOSCALE.md.
Cost-efficiency leaderboard: benchmarks/COST_EFFICIENCY.md.
Strict evidence readiness runbook: benchmarks/STRICT_EVIDENCE_READINESS.md.
Production evidence environment contract: benchmarks/PRODUCTION_EVIDENCE_ENV.md.
Living HTML dashboard: docs/benchmark-dashboard.html.
Machine-readable dashboard status: docs/data/leaderboard-status.json.
The weekly workflow also publishes the refreshed dashboard to GitHub Pages at
caspiang.github.io/wavemind without
writing scheduled bot commits to main.
The status JSON exposes first-class publication_contract, freshness_gate,
agent_quality, agent_impact, structured_memory, memory_os_intelligence,
cluster_autoscale, memory_os_policy, memory_os_policy_evolution,
memory_os_policy_bundle,
production_evidence_env, production_evidence_dispatch,
strict_evidence_readiness, and cost_efficiency
sections, so dashboards can verify the weekly GitHub Pages publication path,
detect stale or missing public evidence, track task success, stale-error
suppression, context savings, active Memory OS worker behavior, policy
decisions, cluster autoscale/operator coverage, and the exact strict-evidence
workflow dispatch plus promotion contract without scraping Markdown.
Production readiness gate: benchmarks/PRODUCTION_READINESS.md
from benchmarks/production_readiness_results.json.
Strict production evidence gate: benchmarks/PRODUCTION_EVIDENCE.md
from benchmarks/production_evidence_results.json. This is the hard boundary
for remote multi-region, managed-serverless, service-backed 10M, and 100M
scale claims. The persisted 50M FAISS profile now passes this gate.
Production evidence environment contract:
benchmarks/PRODUCTION_EVIDENCE_ENV.md
from benchmarks/production_evidence_env_contract.json. This maps every
strict-evidence variable to the claims, workflows, artifacts, GitHub Actions
secrets, workflow inputs, and safe .env.example placeholders without
serializing credential values. Production evidence dispatch plan:
benchmarks/PRODUCTION_EVIDENCE_DISPATCH.md
from benchmarks/production_evidence_dispatch_results.json. This turns the
strict evidence gaps into secret-safe gh workflow run ... payloads, download
commands, and ingest commands for maintainer-reviewed production runs.
Strict evidence readiness runbook:
benchmarks/STRICT_EVIDENCE_READINESS.md
from benchmarks/strict_evidence_readiness_results.json. This joins strict
evidence, preflight, dispatch, scale plans, scale gaps, release claims, and
leaderboard freshness into one operator checklist. It verifies that every
remaining remote/10M-service/100M claim has exact safe dispatch, download, ingest,
strict validation, and refresh commands, while keeping claims locked until real
artifacts pass.
Operator evidence bundle: benchmarks/PRODUCTION_EVIDENCE_BUNDLE.md
from benchmarks/production_evidence_bundle_results.json. This combines the
strict gate, preflight, readiness, artifact audit, claim boundaries, and exact
next actions into one publishable status contract.
Large-N service plans include resumable --checkpoint-path commands so
interrupted 10M/50M/100M ingest runs can continue from completed batches instead
of restarting from zero.
Manual strict-evidence runners include .github/workflows/production-streaming-load.yml,
.github/workflows/external-http-cluster-load.yml,
.github/workflows/external-http-active-active.yml, and
.github/workflows/serverless-observed-telemetry.yml. They run checkpointed
Qdrant, sharded Qdrant, pgvector, FAISS IVF-PQ, remote API-node, remote
active-active, or serverless telemetry profiles on sized infrastructure. The
preferred review path is to leave commit_results=false, download the Actions
artifact, and ingest it locally with the strict artifact gate:
gh run download RUN_ID --name production-streaming-load-results --dir state/large-run
wavemind ingest-production-evidence --artifact-dir state/large-run --refresh
python -m pytest tests/test_production_evidence_ingest.py tests/test_production_evidence_gate.py -qThe ingest command accepts only strict production evidence filenames:
http_cluster_load_results.json, external_http_active_active_results.json,
observed-telemetry.remote.json, or real large-N proof filenames such as
production_streaming_load_qdrant_sharded_100m_results.json. It rejects smoke
artifacts, local/loopback active-active runs, sample endpoints, wrong engines,
wrong vector counts, skipped rows, recall below 0.95, p99 above 100 ms, and
failed SLO/cost rows. This prevents a local transport smoke from accidentally
unlocking a remote Kubernetes/serverless production claim.
Prerequisite preflight: wavemind production-evidence-preflight --write-artifacts.
This writes benchmarks/production_evidence_preflight_results.json and
benchmarks/PRODUCTION_EVIDENCE_PREFLIGHT.md, checks the required remote URLs,
service index env vars, FAISS paths, plan artifacts, disk headroom, and exact
large-run commands, and can fail deployments with
--fail-on-action-required. Operator env contract: wavemind production-evidence-env --write-artifacts. This writes
benchmarks/production_evidence_env_contract.json,
benchmarks/PRODUCTION_EVIDENCE_ENV.md, and
deploy/cluster/production-evidence.env.example, and can fail staging setup
with --fail-on-missing. Strict claim gate: wavemind production-evidence --strict. Dispatch contract: wavemind production-evidence-dispatch --write-artifacts, or wavemind production-evidence-dispatch --fail-on-action-required when the environment must already be ready to launch
all unfinished strict-evidence jobs. Combined operator bundle:
wavemind production-evidence-bundle --write-artifacts, or wavemind production-evidence-bundle --strict when a release must fail unless all
remote/large-N production claims are unlocked. Deployment admission:
wavemind production-admission --target-memories 100000000 --engine qdrant-sharded-service --fail-on-blocked is the final deploy-facing check; it
keeps a requested production profile blocked until its matching strict evidence
artifact passes. Persisted 50M FAISS is admitted; the unfinished 10M service and
100M sharded profiles remain blocked. --allow-plan-only reports the next run
contract without admitting production. The same gate can protect the API process itself:
wavemind serve --require-production-admission --production-target-memories 100000000 --production-engine qdrant-sharded-service. Environment-driven
deployments can set WAVEMIND_REQUIRE_PRODUCTION_ADMISSION=1,
WAVEMIND_PRODUCTION_TARGET_MEMORIES, WAVEMIND_PRODUCTION_ENGINE, and
WAVEMIND_PRODUCTION_ADMISSION_ROOT; the server exits before binding a port
when the requested scale is not admitted.
Weekly benchmark refresh: .github/workflows/benchmark-leaderboard.yml reruns
the fast benchmark profiles, regenerates the benchmark matrix/report/leaderboard
docs/assets/benchmark-summary.svg, docs/benchmark-dashboard.html, the
agent-impact leaderboard, the structured-memory report, the
Memory OS intelligence report, the cluster-autoscale report,
cost-efficiency leaderboard,
production-readiness report, the strict production-evidence report, the
production-evidence dispatch plan, the combined production-evidence bundle, and
the strict evidence readiness runbook, and the production-admission report,
validates freshness with benchmarks/validate_benchmark_artifacts.py, writes
benchmarks/benchmark_artifact_audit.json, renders
docs/data/leaderboard-status.json, and uploads changed benchmark artifacts for
maintainer review. It also builds a static GitHub Pages bundle from the
dashboard, reports, summary SVG, public status JSON, and machine-readable JSON evidence,
then deploys the living leaderboard with actions/upload-pages-artifact@v3 and
actions/deploy-pages@v4. docs/data/leaderboard-status.json records this as a
machine-readable publication_contract, including the cron schedule, Pages
deployment actions, status JSON path, review policy, and claim boundary. The
workflow does not push scheduled bot commits to main; reviewed benchmark
refreshes should be committed from a maintainer account.
full-check and the release workflow also run the same freshness gate with
--max-age-days 8, so stale or manually edited public benchmark artifacts block
normal CI and package releases.
External cluster benchmark refresh: .github/workflows/external-http-cluster-load.yml
runs benchmarks/http_cluster_load_benchmark.py against real API-node URLs or a
JSON node manifest, including the external /query/batch recall check, and can
commit benchmarks/http_cluster_load_results.json plus refreshed leaderboard
artifacts when commit_results=true. The four-node kind workflow also binds this
load profile to a physical worker outage over pod DNS and verifies target-specific
cluster admission before publishing evidence.
External active-active refresh: .github/workflows/external-http-active-active.yml
runs benchmarks/local_http_active_active_smoke.py against real API-region URLs
or a JSON region manifest and can commit
benchmarks/external_http_active_active_results.json plus refreshed
leaderboard/readiness artifacts when commit_results=true.
Remote serverless telemetry refresh: .github/workflows/serverless-observed-telemetry.yml
runs benchmarks/serverless_observed_telemetry_benchmark.py against deployed
HTTP/HTTPS API node URLs, uploads deploy/serverless/observed-telemetry.remote.json,
and can commit the remote telemetry plus refreshed leaderboard/readiness
artifacts when commit_results=true. When that remote file exists,
benchmarks/scale_readiness_benchmark.py prefers it over the checked-in
loopback telemetry.
The compact leaderboard now carries an explicit evidence-status table:
benchmarks/BENCHMARK_LEADERBOARD.md.
Use that generated file for exact current numbers. This README keeps the
public claim boundaries stable:
| Claim area | Current public status | Source of truth | Not proven yet |
|---|---|---|---|
| Production readiness | WaveMind core readiness is gated by checked-in artifacts before release. | benchmarks/production_readiness_results.json, benchmarks/PRODUCTION_READINESS.md |
Missing external competitor credentials should not be treated as WaveMind core failure, but they still limit competitor claims. |
| Strict production evidence | The gate now passes 2/8 requirements: persisted 50M FAISS is proven at recall@10 0.9705 and p99 73.11 ms; the non-loopback Kubernetes cluster passes success/failover 1.00, query p99 84.80 ms, batch p99 148.82 ms, and physical-worker attestation 10/10. |
benchmarks/production_evidence_results.json, benchmarks/PRODUCTION_EVIDENCE.md, benchmarks/production_evidence_gate.py, wavemind production-evidence --strict |
Six requirements remain: remote active-active, managed serverless telemetry, three 10M service profiles, and the 100M sharded service run. |
| Production evidence preflight | Remote endpoint/env/path prerequisites are checked before launching expensive strict-evidence jobs. | benchmarks/production_evidence_preflight_results.json, benchmarks/PRODUCTION_EVIDENCE_PREFLIGHT.md, wavemind production-evidence-preflight --write-artifacts |
A ready preflight is not a passing evidence result; it only proves the environment is ready to run the remote/large-N jobs. |
| Production evidence env contract | Secret-safe operator map from every strict-evidence env var to the workflows, claims, artifacts, GitHub Actions secrets, input bindings, and .env.example placeholders it unlocks. |
benchmarks/production_evidence_env_contract.json, benchmarks/PRODUCTION_EVIDENCE_ENV.md, deploy/cluster/production-evidence.env.example, wavemind production-evidence-env --write-artifacts |
It does not unlock production claims; it prevents ambiguous or unsafe production evidence launches and keeps secrets out of checked-in artifacts. |
| Production evidence dispatch | Secret-safe workflow dispatch contract for every unfinished strict-evidence job, including safe commit_results=false launch commands, publish commands, required env/secrets, and artifact promotion commands. |
benchmarks/production_evidence_dispatch_results.json, benchmarks/PRODUCTION_EVIDENCE_DISPATCH.md, wavemind production-evidence-dispatch --write-artifacts |
A dispatch plan only launches or reviews evidence runs; it does not unlock production claims until downloaded artifacts pass ingest and strict validation. |
| Strict evidence readiness | Operator runbook that joins strict evidence, preflight, dispatch, scale plans, scale gaps, release claims, and freshness into one table of blockers, locked claims, safe dispatch commands, ingest commands, and validation commands. | benchmarks/strict_evidence_readiness_results.json, benchmarks/STRICT_EVIDENCE_READINESS.md, python benchmarks/strict_evidence_readiness_report.py |
Current readiness is action_required: non-loopback Kubernetes cluster evidence and 50M FAISS are complete; the remaining six remote/service jobs are mapped but not yet proven. |
| Production evidence bundle | Single operator-facing status contract that combines strict gate, preflight, readiness, artifact audit, claim boundaries, next actions, and release exit behavior. | benchmarks/production_evidence_bundle_results.json, benchmarks/PRODUCTION_EVIDENCE_BUNDLE.md, wavemind production-evidence-bundle --write-artifacts |
claims_limited is expected until the strict remote/large-N artifacts pass. |
| Release claims | Compact release-facing claim contract for GitHub Releases and launch posts: what is safe to claim, what remains locked, and which command unlocks the next evidence tier. | benchmarks/release_claims_results.json, benchmarks/RELEASE_CLAIMS.md, wavemind release-claims --write-artifacts --fail-on-blocked |
core_release_ready, non-loopback Kubernetes cluster SLO, and the 50M persisted FAISS profile are supported; remote multi-region and 10M-100M service claims remain locked. |
| Agent impact leaderboard | Behavioral benchmark evidence is aggregated across agent coherence, dynamic-memory policy, long-memory retrieval, and LongMemEval answer quality. | benchmarks/agent_impact_results.json, benchmarks/AGENT_IMPACT.md, benchmarks/agent_impact_leaderboard.py |
It proves lift on the listed checked-in scenarios only; it does not claim general agent success outside those tasks. |
| Memory OS intelligence | Adaptive-worker evidence is aggregated across scale readiness, agent coherence, staging canary, and admission artifacts. It tracks hot-query prewarm, transition-learned predictive prefetch, priority learning, adaptive forgetting, concept consolidation, Redis coordination, canary status, and production-admission boundaries. | benchmarks/memory_os_intelligence_results.json, benchmarks/MEMORY_OS_INTELLIGENCE.md, benchmarks/memory_os_intelligence_report.py |
It proves Memory OS behavior on checked-in fixtures; unattended production automation remains locked until real shared Redis, distributed lock, runtime env, and large-scale evidence pass. |
| Cluster autoscale | Cluster/operator evidence is pulled into a dedicated public report. It tracks deterministic shard placement, node/zone loss availability, autoscale planning, rebalance checkpoints, Kubernetes operator reconciliation, quorum safety, HTTP sharding, active-active convergence, CRDT field state, and the 100M capacity envelope. | benchmarks/cluster_autoscale_results.json, benchmarks/CLUSTER_AUTOSCALE.md, benchmarks/cluster_autoscale_report.py |
It is a deterministic capacity and operator evidence report, not a real 100M vector-query latency benchmark or managed Kubernetes production SLO. |
| Kubernetes physical worker failure | Four WaveMind pod-DNS endpoints across three kind worker zones retain 1.00 recall while one worker container is physically paused, recover without pod replacement, then pass the mixed cluster load at query p99 84.80 ms and batch p99 148.82 ms. |
benchmarks/kubernetes_cluster_network_smoke_results.json, benchmarks/http_cluster_load_results.json, benchmarks/kubernetes_cluster_network_smoke.py, workflow run 29070578441 |
This unlocks the non-loopback Kubernetes service-node SLO. It does not claim managed Kubernetes, independent remote regions, or 10M-100M distributed scale. |
| Kubernetes active-active region failure | Three PVC-backed replicated regions in three worker zones converge 48 initial writes, continue 32 writes plus a delete while region B is physically unavailable, then recover at 1.00 convergence and delete suppression with an idempotent final sync. |
benchmarks/kubernetes_active_active_region_smoke_results.json, benchmarks/kubernetes_active_active_region_smoke.py, wavemind active-active-drill |
This proves the active-active protocol across non-loopback Kubernetes services and a physical zone outage in ephemeral CI. Independent remote regions are still required for strict active-active admission. |
| Kubernetes serverless lifecycle | PVC-backed PostgreSQL, Qdrant, and Redis preserve 24/24 memories through two scale-to-zero cycles; three zone-spread workers achieve write/delete coherence at 3/3 within 1.14 s, and burst p99 remains below 2 s. |
benchmarks/kubernetes_serverless_lifecycle_smoke_results.json, benchmarks/kubernetes_serverless_lifecycle_smoke.py, .github/workflows/kubernetes-operator-smoke.yml |
This proves external-state lifecycle and bounded worker-cache convergence in ephemeral non-loopback Kubernetes. Managed Knative/KEDA endpoints and remote telemetry are still required for strict serverless admission. |
| Kubernetes PostgreSQL/Qdrant DR | A checksummed PostgreSQL backup restores into an independent namespace with fresh PVCs and an empty Qdrant service; recall and index parity are 24/24, including after recovery API replacement. |
benchmarks/kubernetes_postgres_qdrant_dr_smoke_results.json, benchmarks/kubernetes_postgres_qdrant_dr_smoke.py, .github/workflows/kubernetes-operator-smoke.yml |
This proves logical backup/restore and vector-index reconstruction in ephemeral Kubernetes. It is not managed PostgreSQL PITR, remote object-store recovery, or multi-region DR. |
| Scale gap matrix | Large-N proof status for 10M Qdrant, 10M sharded Qdrant, 10M pgvector, 50M FAISS IVF-PQ, and 100M sharded Qdrant. It joins strict evidence, preflight, run commands, missing env, and measured baselines. | benchmarks/scale_gap_results.json, benchmarks/SCALE_GAP.md, wavemind scale-gap --write-artifacts |
1/5 profiles are complete with 50,000,000 proven vectors; the three 10M service profiles and 100M sharded Qdrant remain. |
| Cost-efficiency leaderboard | Cost, latency, recall, SLO, and memory-count evidence are ranked across measured production-load artifacts and the remaining plan-only 10M-service/100M contracts. | benchmarks/cost_efficiency_results.json, benchmarks/COST_EFFICIENCY.md, benchmarks/cost_efficiency_leaderboard.py |
The measured 50M row is evidence; remaining planned rows are capacity/cost contracts only. |
| Production admission | Deployment-facing gate for a requested memory count and engine. It maps the requested 10M/50M/100M deployment to the required strict evidence profile and fails deploys until that artifact passes. | benchmarks/production_admission_results.json, benchmarks/PRODUCTION_ADMISSION.md, wavemind production-admission --target-memories 100000000 --engine qdrant-sharded-service --fail-on-blocked |
Current 100M sharded Qdrant status is plan_only, not admitted: the run contract exists, but production_streaming_load_qdrant_sharded_100m_results.json is still missing. |
| Cluster admission | Deployment-facing gate for non-loopback service-node rollouts. It requires strict load evidence, a ready preflight, and an exact node ID-to-URL match for the requested target. | benchmarks/cluster_admission_results.json, benchmarks/CLUSTER_ADMISSION.md, wavemind cluster-admission --fail-on-blocked --write-artifacts |
The attested kind target is admitted. A different staging or production target remains blocked until it produces matching endpoint-specific evidence. |
| Active-active admission | Deployment-facing gate for remote multi-region active-active rollout. It admits only when the strict external HTTP active-active artifact passes; local/loopback runs remain development evidence. | benchmarks/active_active_admission_results.json, benchmarks/ACTIVE_ACTIVE_ADMISSION.md, wavemind active-active-admission --allow-plan-only --write-artifacts |
Current status is plan_only, not admitted: benchmarks/external_http_active_active_results.json is still missing and remote region env is not configured. |
| Serverless admission | Deployment-facing gate for managed/serverless rollout. It admits only when remote deployed API nodes produce strict telemetry; loopback telemetry remains development evidence. | benchmarks/serverless_admission_results.json, benchmarks/SERVERLESS_ADMISSION.md, wavemind serverless-admission --allow-plan-only --write-artifacts |
Current status is plan_only, not admitted: deploy/serverless/observed-telemetry.remote.json is still missing and WAVEMIND_SERVERLESS_NODES is not configured. |
| External multimodal evidence runner | Reproducible path from a real external encoder/object-store manifest to the admission artifact. It validates external vectors, target queries, s3:// asset URIs, object-store verification metadata, vector persistence, provenance, routing, precision, query latency, and encode p95 fields. |
wavemind multimodal-external-evidence --manifest external_multimodal_manifest.json --write-artifacts --output benchmarks/multimodal_external_encoder_results.json |
Runner-ready. No checked production artifact is included until a real external manifest is available. |
| Multimodal admission | Deployment-facing gate for production multimodal memory claims. It admits only when the structured-memory contract passes and a real external encoder/object-store artifact satisfies modality count, payload/query volume, precision, cross-modal routing, vector persistence, provenance, p99 query latency, encode p95, and error-rate thresholds. | benchmarks/multimodal_admission_results.json, benchmarks/MULTIMODAL_ADMISSION.md, wavemind multimodal-admission --allow-plan-only --write-artifacts |
Current status is plan_only, not admitted: deterministic structured-memory evidence passes, but benchmarks/multimodal_external_encoder_results.json is still missing. |
| Memory OS canary | Staging proof that representative query-audit traffic can drive Memory OS prewarm, predictive prefetch, priority learning, TTL cleanup, and admission. | benchmarks/memory_os_canary_results.json, benchmarks/MEMORY_OS_CANARY.md, wavemind memory-os-canary --target-memories 100000 --namespace-count 64 --deployment staging --write-artifacts |
This is not remote Kubernetes, real Redis, or 10M/100M production evidence; it only proves the worker/admission contract under seeded staging traffic. |
| Memory OS policy evolution | Multi-cycle Memory OS proof that repeated policy gaps are remembered and influence later scheduler plans. It verifies full policy coverage, repeated required-policy escalation, stable OK policy detection, hot-query prewarm, predictive prefetch, priority learning, and required worker task coverage. | benchmarks/memory_os_policy_evolution_results.json, benchmarks/MEMORY_OS_POLICY_EVOLUTION.md, wavemind memory-os-evolution --cycles 3 --write-artifacts |
Current status is pass on deterministic local/staging evidence. It does not unlock unattended production automation without remote Redis, distributed lock, runtime env, and strict large-scale evidence. |
| Memory OS policy bundle | Operator-facing runtime policy manifest generated from canary, policy-evolution, and admission artifacts. It emits enabled task ids, required Redis/lock env, observability metrics, Kubernetes/CronJob patch data, and promotion gates. | benchmarks/memory_os_policy_bundle_results.json, benchmarks/MEMORY_OS_POLICY_BUNDLE.md, wavemind memory-os-policy-bundle --write-artifacts |
Current status is staging_ready: it can promote checked Memory OS behavior to staging, but production remains locked while memory-os-admission is plan_only. |
| Memory OS admission | Deployment-facing gate for adaptive workers. It checks scheduler safety, hot-query audit signal, Redis cache wiring, distributed lock wiring, singleton/idempotent mutations, policy coverage, and strict architecture boundaries before Memory OS workers become production automation. | benchmarks/memory_os_admission_results.json, benchmarks/MEMORY_OS_ADMISSION.md, wavemind memory-os-admission --target-memories 10000000 --namespace-count 4096 --deployment production --allow-plan-only --write-artifacts |
Current 10M Memory OS status is plan_only, not admitted: the worker plan exists, but staging query-audit traffic, Redis/lock runtime env, and strict million-plus architecture evidence are still required. |
| Production scale run planner | One command plans the next large-N jobs across 10M Qdrant, 10M sharded Qdrant, 10M pgvector, 50M FAISS IVF-PQ, and 100M sharded Qdrant, including env, checkpoint, storage, SLO, monthly budget, cost per 1M memories, compute cost per 1M queries, plan-only Pareto frontier, and output artifact contracts. | benchmarks/production_scale_run_plan.json, wavemind production-scale-plan --write-artifact |
This is a run contract and preflight only; it does not replace the real latency/recall result artifacts. |
| 10M memory-scale profile | Checked-in compressed FAISS IVF-PQ streaming profile exists and is reported in the generated leaderboard. | benchmarks/production_streaming_load_ivfpq_10m_results.json |
Not yet a completed 10M Qdrant/pgvector service comparison. |
| 50M persisted FAISS IVF-PQ | Real GitHub-hosted run over 50,000,000 128D vectors and 2,000 queries. Adaptive nprobe selected 512: recall@10 0.9705, p99 73.11 ms, valid cost/SLO evidence. |
benchmarks/production_streaming_load_ivfpq_50m_results.json, benchmarks/production_streaming_load_50m_plan.json |
This proves a compressed single-node persisted index, not a 100M distributed service cluster. |
| pgvector tuning | Real PostgreSQL/pgvector service profile now separates baseline HNSW, exact recall floor, and iterative HNSW tuning. | benchmarks/production_pgvector_tuning_results.json |
This is a 50k service-backed tuning profile, not yet the 100k/1M production load SLO artifact. |
| Qdrant streaming | Real Qdrant streaming smoke exists, a tuned 1M service run passes SLO after warmup/chunking, a real two-service sharded Qdrant smoke passes, and single-service plus horizontally sharded 10M/100M plan-only contracts are checked in. | benchmarks/production_streaming_load_qdrant_smoke_results.json, benchmarks/production_streaming_load_qdrant_1m_results.json, benchmarks/production_streaming_load_qdrant_1m_tuned_results.json, benchmarks/production_streaming_load_qdrant_sharded_smoke_results.json, benchmarks/production_streaming_load_qdrant_10m_plan.json, benchmarks/production_streaming_load_qdrant_sharded_10m_plan.json, benchmarks/production_streaming_load_qdrant_sharded_100m_plan.json |
The 10M/100M Qdrant service results are not claimed until the matching result artifacts are produced by real runs. |
| pgvector streaming | Real PostgreSQL/pgvector streaming smoke exists, and a 10M plan-only service contract is checked in. | benchmarks/production_streaming_load_pgvector_smoke_results.json, benchmarks/production_streaming_load_pgvector_10m_plan.json |
The 10M pgvector service result is not claimed until production_streaming_load_pgvector_10m_results.json is produced by a real run. |
| HTTP cluster load | The checked artifact runs the mixed workload from inside Kubernetes against four pod-DNS API nodes: success/failover/delete suppression 1.00, repaired replica 1, query p99 84.80 ms, batch p99 148.82 ms, and external batch requests 24 -> 1. Bulk lifecycle batch p99 is reported separately at 6694.76 ms. |
benchmarks/http_cluster_load_results.json, .github/workflows/kubernetes-operator-smoke.yml, .github/workflows/external-http-cluster-load.yml |
This is ephemeral non-loopback Kubernetes evidence, not managed multi-region or million-scale service evidence. |
| HTTP active-active regions | Local multi-process API-region evidence exists, and the external URL-based contract now has a loopback artifact. The external workflow can run the same namespace-delta contract against real regional API URLs. | benchmarks/local_http_active_active_smoke_results.json, benchmarks/external_http_active_active_loopback_results.json, .github/workflows/external-http-active-active.yml |
Local/loopback active-active evidence is not a remote Kubernetes/serverless multi-region result until benchmarks/external_http_active_active_results.json is produced by a real run. |
| Serverless telemetry | Loopback replica telemetry exists; remote telemetry has a dedicated manual workflow and artifact path. | deploy/serverless/observed-telemetry.loopback.json, .github/workflows/serverless-observed-telemetry.yml |
Loopback evidence is not a hosted managed-serverless claim until observed-telemetry.remote.json is committed. |
| Competitor adapters | Local Mem0/LangGraph/GraphRAG-style adapters run; optional Zep evidence is skipped until configured. | benchmarks/memory_competitor_results.json |
Not a full independent Mem0/Zep/Letta leaderboard without live service credentials and public runner parity. |
Visual summary generated from the checked-in JSON results:
Regenerate the matrix and chart locally:
python benchmarks/benchmark_registry.py --output benchmarks/benchmark_matrix_results.json
python benchmarks/render_benchmark_charts.py --output docs/assets/benchmark-summary.svgThe chart shows completed local measurements plus the public benchmark roadmap. Planned public benchmarks stay out of the results section until the dataset, engine, and result JSON are committed.
Status legend:
implemented- script and checked-in result exist.runner ready- adapter exists, but the official public dataset result is not checked in yet.planned- benchmark is part of the public proof path, but no WaveMind result is claimed.
How to read the benchmark classes:
| class | Popular examples | What it answers for WaveMind |
|---|---|---|
| Retrieval / embeddings | BEIR, MTEB Retrieval, MIRACL | Does WaveMind preserve normal vector-search quality on public qrels? |
| Vector index / database | ANN-Benchmarks, VectorDBBench | Is the candidate index fast enough at scale? |
| Agent memory | LoCoMo, LongMemEval, LongMemEval-V2, LMEB | Does WaveMind retrieve the right evolving memory across long histories? |
| RAG quality | RAGBench | Does dynamic memory improve final context and answer quality? |
Current read:
| area | result | honest interpretation |
|---|---|---|
| Public agent-memory evidence | On official LoCoMo locomo10.json, WaveMind reaches evidence_recall@5 0.386 with hash embeddings and 0.547 with sentence-transformers. Fair namespace-filtered Chroma reaches 0.257 / 0.407; Qdrant reaches 0.263 / 0.409. |
WaveMind retrieves more labeled evidence. Chroma is still the fastest static vector-store baseline. Qdrant local payload filtering is much slower than service-mode Qdrant should be. |
| Public retrieval sanity check | On BEIR SciFact, WaveMind reaches nDCG@10 0.354, Recall@10 0.482; Qdrant matches that quality; Chroma reaches 0.350 / 0.467 with identical hash embeddings. |
Same-embedding retrieval quality is close. Chroma is fastest at 1.79 ms; Qdrant local is 17.71 ms; WaveMind exact path is 117.02 ms. |
| Public multilingual retrieval | On NoMIRACL Russian, sampled at 200 queries / 5000 compact candidate passages, WaveMind reaches nDCG@10 0.434, Recall@10 0.516, matching Qdrant and staying within 0.002 nDCG of Chroma on identical hash embeddings. |
Russian same-embedding quality is at parity. Chroma is faster at 2.60 ms; WaveMind is 10.22 ms; Qdrant local is 18.86 ms. |
| Static agent recall | WaveMind precision@1 equals Chroma at 0.82; WaveMind precision@3 is 0.90 vs Chroma 0.88. |
Competitive quality, but Chroma is faster on the static vector-store path. |
| Agent coherence and token savings | On a 500-memory long user-history task simulation, WaveMind reaches 0.92 task success, 0.00 stale error rate, 0.93 context saved, 9 coherent turns, and 2.65 ms average query latency. WaveMind + Memory OS keeps 0.92 task success while proving 1 hot query, 1 prewarmed query, 4 predictive-prefetch warmed queries, 5 priority predictions, and cache hit rate 0.24. Static vector reaches 0.33 task success and 0.73 stale error rate; Chroma static reaches 0.33 task success and 0.45 stale error rate. |
This measures agent behavior rather than nearest-neighbor retrieval: fewer stale facts, longer coherent task runs, compact context, and observable Memory OS learning signals. |
| Dynamic memory policy | WaveMind reaches 1.00 stale suppression; Chroma static is 0.00. |
This is the strongest current differentiation: hotness, TTL, corrections, and namespaces. |
| Field memory dynamics | Graph-enabled WaveMind reaches 1.00 precision@1, 1.00 stale suppression, 1.00 concept formation, and 1.00 durable concept consolidation vs static WaveMind at 0.20 / 0.20 / 0.00 / 0.00. |
This is still synthetic, but it is now a regression check for memory-to-memory excitation, conflict inhibition, decay, and self-organization into auditable concept memories. |
| Long-term evidence | WaveMind reaches 1.00 evidence recall@5, 1.00 precision@1, and 1.00 stale suppression on the synthetic long-memory evidence benchmark. |
This is the first proof-shaped benchmark for agent memory: it measures whether stale/corrected/expired/cross-user facts stay out of retrieved evidence. |
| Capacity | Static precision@1 is 0.94 at 5000 memories; dynamic policy keeps 1.00 on the current checks. |
Quality is holding on these checks, but dynamic latency must be optimized. |
| LongMemEval full retrieval | On the official LongMemEval-S cleaned file, 470 non-abstention session-level questions, WaveMind reaches evidence_recall@5 0.782 and precision@1 0.696; Chroma static reaches 0.518 / 0.355; Qdrant static reaches 0.520 / 0.355. |
This is now the strongest public memory result in the repo. It is retrieval-only, not final answer quality. |
| LongMemEval 50-query smoke | On the first 50 non-abstention LongMemEval-S questions, WaveMind reaches evidence_recall@5 0.920, precision@1 0.760, and MRR@5 0.827; Chroma/Qdrant static reach 0.600, 0.260, and 0.385. |
This is the fast regression profile for checking current changes before rerunning the full LongMemEval profile. WaveMind wins on quality; latency still needs work. |
| ANN/index curve | At 50000 generated 128-d vectors, NumPy exact keeps recall@10 1.000 at 1.99 ms; quantized int8 keeps 0.934 at 16.27 ms; Annoy is faster at 3.21 ms but drops to 0.730 recall; FAISS flat keeps 1.000 recall with a cold-start-inflated 80.47 ms average and 2.42 ms p95; Qdrant local keeps 1.000 recall at 33.82 ms. |
Current local scale boundary is clear: top-k selection is faster, quantized is memory-oriented but still not a latency win, Annoy needs recall tuning, and service-mode indexes are the production path. |
| Production load | At 100000 generated 128-d vectors, service-mode Qdrant reaches recall@10 1.000, avg 10.28 ms, p99 21.26 ms, passes the checked-in production SLO gate (recall >= 0.95, p99 <= 100 ms, 100 qps, 3 replicas, HPA max 24), and estimates $1.39 per 1M queries with $365.02 monthly target cost. At 1M over 100 queries, persisted FAISS reaches recall@10 1.000, avg 39.12 ms, p99 57.71 ms, and estimates $4.17 per 1M queries with 6 replicas for 100 qps. The older 1M Qdrant tuned production-load profile reaches recall@10 0.984, avg 82.57 ms, p99 137.86 ms; the newer streaming 1M Qdrant profile closes that p99 gap after safe upsert chunking, wait-after-build, and query warmup. |
100k Qdrant, 1M persisted FAISS, and the tuned 1M Qdrant streaming profile now pass recall/p99 production gates on the tested machine. The older Qdrant load profile stays checked in as evidence that cold/untuned service tails can fail. |
| pgvector tuning | On a real PostgreSQL/pgvector service at 50000 vectors, baseline HNSW reaches recall@10 0.834, exact mode reaches 1.000 with p99 76.98 ms, and iterative HNSW reaches 0.970 with p99 55.19 ms. Qdrant service remains the speed reference at 1.000 recall and p99 17.84 ms. |
pgvector now has a measured production tuning path. Exact mode proves correctness; iterative scan passes the 50k recall/p99 gate and should be promoted into 100k/1M load profiles next. |
| Streaming production load | benchmarks/production_streaming_load_benchmark.py generates and inserts vectors in batches, stores only query source vectors outside the index, and measures target-recall, p99, SLO, and cost. The checked-in 10M compressed FAISS IVF-PQ run reaches recall@10 0.990, p99 60.13 ms; the real 50M persisted FAISS run reaches recall@10 0.9705, p99 73.11 ms over 2,000 queries after an adaptive nprobe sweep; tuned 1M Qdrant reaches recall@10 1.000, p99 26.37 ms; service smokes cover Qdrant, two-service sharded Qdrant, and pgvector. |
The 10M and 50M FAISS results are compressed target-recall profiles, not exact-neighbor recall. The unfinished strict profiles are 10M Qdrant, 10M sharded Qdrant, 10M pgvector, and 100M sharded Qdrant. |
| Structured memory report | Dedicated status view for image/audio/video/3D/table/event/graph payloads. It reports 7 modalities, structured precision@1 1.000, cross-modal precision@1 1.000, persisted vector rate 1.000, provenance 1.000, precomputed-vector precision@1 1.000, temporal event precision@1 1.000, knowledge-graph precision@1 1.000, graph path precision@1 1.000, and all gate checks passing. |
This makes the multimodal/structured roadmap evidence visible outside the larger scale-readiness row. It is still a deterministic fixture and external-vector contract, not a claim of broad production multimodal model quality. |
| Scale readiness | Deterministic 1M-memory simulation validates 4096 namespace placements over 4 nodes with replication factor 2, node-loss availability 1.000, zone-loss availability 1.000, Kubernetes StatefulSet, HorizontalPodAutoscaler, repair CronJob, Memory OS CronJob, majority control-plane lease/config revision safety with stale-leader, stale-revision, and minority-commit rejection, operator-style WaveMindCluster reconciliation for 4096 namespaces, operator status phase Ready with resources/capacity/autoscaling/repair/Memory OS/production-admission/control-plane conditions true, hot-cache hit rate 0.920, query-audit prewarm warmed 1 query with prewarm hit true, query-vector cache local hit rate 0.995 with 1 encode call, Redis query-vector cache shared across workers true, FastAPI /query/batch answers 100 recall queries with 1 HTTP request and batch hit rate 0.990, shared Redis rate limiter allows 4 and limits 1 across 2 workers, Redis-compatible shared cache visible across workers, Memory OS Redis prewarm warmed 2 queries, predictive prefetch warmed 6 queries, transition-prefetch hit true on budget recall -> risk limits, explicit feedback events 8, Redis Memory OS demoted cold memories, Redis cross-worker hit true, Redis namespace invalidation true, Redis Memory OS architecture advice architecture_required, API cache invalidation on /remember, /feedback, /feedback/batch, and /forget prevents stale cached recall, batch feedback accepts 2, rejects 1, writes 2 audit events, and updates positive/negative priority, Memory OS found 2 hot queries, warmed 2, predictively warmed 6, demoted cold memories, purged 1 expired memory, created 1 durable concept, emitted a production execution plan with safe_to_run=true, Redis and lock environment requirements, singleton state-mutating tasks, and worker-pool cache-prewarm, and emitted production architecture advice for service-index, namespace-sharding, production-controls, replication-capacity, load-test, and multimodal readiness, service-mode distributed sharding recall after primary loss, service-mode repair copied 1 missing replica record with recall after repair true, real HTTP shard transport with proxy bypass true, HTTP repair copied 1 missing replica record, HTTP tombstone repair deleted 1 stale API record, concurrent HTTP writes 12, concurrent query hit rate 1.000, service-mode tombstone suppression before repair true, tombstone repair deleted 1 stale replica record, suppression after repair true, anti-entropy worker repaired 1 missing record and deleted 1 stale tombstone record, quorum-replicated runtime recall after node loss, missing-record repair, tombstone repair, concurrent runtime writes 12, concurrent runtime query hit rate 1.000, active-active namespace delta sync with cursor-based incremental record export and field-only hotness export, sustained active-active mesh sync across 3 independent regions and 3 namespaces with convergence 1.000, delete suppression 1.000, success 1.000, 90 pair syncs, final no-op imports 0, HTTP service-region active-active sync through FastAPI delta endpoints with convergence 1.000 and final no-op imports 0, real local HTTP active-active service-region smoke with convergence 1.000, delete suppression 1.000, success 1.000, and final no-op imports 0, field-state CRDT convergence/idempotency/tombstone-wins, checksummed replicated snapshot/restore, offsite mirror verification, portable archive verification, S3-compatible upload/latest-metadata/download/retention verification, object-store DR drill true, SQLite recovery journal full restore true, point-in-time restore true, structured payload precision@1 1.000, cross-modal target-modality precision@1 1.000, cross-modal vector persistence 1.000, precomputed external-vector precision@1 1.000, precomputed vector persistence 1.000, cross-modal provenance rate 1.000, temporal event precision@1 1.000, temporal persistence 1.000, temporal provenance 1.000, knowledge-graph direct/two-hop/three-hop/predicate precision 1/1/1/1, graph path precision@1 1.000, graph persistence 1.000, graph provenance 1.000, and a deterministic 100M-memory capacity envelope with weighted rendezvous zone-aware placement, 128 nodes, 8 zones, RF=3, node/zone-loss availability 1.000, distinct replica rate 1.000, zone-spread rate 1.000, valid capacity plan true, and a 128-to-160-node scale-out movement audit. |
This proves routing, control-plane split-brain protection for config changes, Kubernetes deployment/operator/HPA/repair/Memory OS manifests and status conditions, production admission wiring for 10M+ targets, service-mode repair, real HTTP shard transport, service-boundary active-active delta sync, real multi-process active-active service-region sync, concurrent API safety for local WaveMind/SQLite nodes, concurrent replicated-runtime safety, tombstone-aware delete repair, anti-entropy background repair, explicit and batch recall feedback, query-vector cache, API batch recall, shared rate limiting, Memory OS adaptive prewarm/transition-learned predictive prefetch/consolidation/forgetting/architecture advice, Memory OS rollout safety contracts, local and Redis-compatible cache prewarm, shared cache invalidation and mutation safety, structured, temporal, cross-modal, and knowledge-graph payload retrieval, external precomputed-vector compatibility, distributed sharding, replicated-runtime, cursor-bounded namespace-delta sync, sustained active-active mesh convergence, distributed field-state convergence, offsite/archive/object-store backup lifecycle, SQLite point-in-time recovery, restore-drill foundations, and 100M-scale placement/capacity planning. |
| Local HTTP cluster smoke | 4 real localhost API processes with isolated SQLite stores, RF=3, read_fanout=1, workers 4: success 1.000, failover hit 1.000, delete suppression 1.000, repaired replicas 1, health true, degraded nodes 0, p99 348.83 ms, SLO true. |
This is the CI-friendly service-mode gate between in-process tests and remote external-node benchmarks. It catches HTTP transport, quorum, repair, delete-suppression, post-load node-health, and circuit-state regressions without needing external infrastructure. |
| Production readiness gate | Current WaveMind core gate score is 1.000: 39/39 criteria pass, 0 require action, 0 fail. Live Zep competitor evidence is tracked separately and remains pending until a real ZEP_API_URL or ZEP_API_KEY is configured. |
This keeps production claims honest without letting a missing commercial competitor credential block WaveMind's own readiness verdict. WaveMind has real 10M and 50M compressed FAISS evidence, measured pgvector tuning, Qdrant/pgvector service smokes, tuned 1M Qdrant SLO evidence, physical Kubernetes failure/DR checks, and a deterministic 100M capacity envelope. |
| Memory competitor adapters | Generated dynamic profile: 50 users, 450 facts, 300 checks. WaveMind reaches precision@1 0.80, precision@3 0.94, stale suppression 1.00. Mem0 runs locally with Qdrant + FastEmbed and reaches 0.68, 0.99, stale suppression 0.60. LangGraph persistent SQLite reaches 0.80, 0.95, stale suppression 1.00. GraphRAG-style static graph reaches 0.85, 0.96, stale suppression 1.00. Zep has live adapter paths for the current zep-cloud Graph API and legacy/OSS-compatible zep-python; it is skipped only until ZEP_API_URL or ZEP_API_KEY points at a real Zep service. |
This prevents fake competitor claims while still checking real installed competitors when they are available. |
| LongMemEval local answer generation | With the same local Ollama qwen2.5:1.5b, WaveMind reaches exact_match 0.240, contains_answer 0.380, token_f1 0.333, and evidence_recall@5 0.920; Chroma and Qdrant static both reach 0.120, 0.160, 0.170, and 0.600. |
This is the first checked-in end-to-end answer benchmark against Chroma/Qdrant. It is still a 50-question lightweight smoke run, not a full LongMemEval leaderboard score. |
The serverless part of scale readiness now includes an operational preflight, not only manifests: target load, required replicas, burst capacity, scale-to-zero safety, cold-start budget, and modeled compute cost are checked by the production gate.
| benchmark | what it proves | status | baseline / competitor | target |
|---|---|---|---|---|
| Agent user-memory retrieval | Natural-language recall over 200 user facts. | implemented | Chroma | Match Chroma precision@1, beat precision@3, stay under 5 ms at 200 memories. |
| Agent coherence and token savings | Simulated agent tasks over long user history: task success, top-1 decision success, stale errors, coherent turns, context saved, and Memory OS learning signals. | implemented | Static vector / Chroma static | Prove WaveMind improves agent behavior, not only nearest-neighbor retrieval, and show that Memory OS produces observable prewarm/prefetch/priority evidence. |
| Dynamic memory policy | Hot memory, TTL, corrections, stale suppression, namespace isolation. | implemented | Chroma static | Keep precision@1 and stale suppression at 1.00, cut avg latency below 10 ms at 1000 memories. |
| Field memory graph dynamics | Related memories excite each other, newer conflicting memories suppress stale facts, graph energy decays, and active clusters can become durable concept memories. | implemented | WaveMind static | Keep precision@1, stale suppression, concept formation, and concept consolidation at 1.00 while moving from synthetic checks to LoCoMo/LongMemEval evidence. |
| WaveMind capacity curve | How recall and latency change at 200 / 1000 / 5000 memories. | implemented | WaveMind-only today | Keep precision@1 >= 0.95 at 5000 memories and dynamic latency below 20 ms. |
| Long-term memory evidence | Evidence retrieval from long histories with profile, preference, correction, TTL, namespace, and filler noise. | implemented | Static vector / Chroma / Qdrant | Keep this as a small regression test while public LoCoMo and LongMemEval runners carry the stronger evidence claims. |
| BEIR-style open retrieval runner | Public corpus.jsonl, queries.jsonl, qrels/*.tsv datasets with the same metrics for each engine. |
implemented | WaveMind / Chroma / Qdrant | Use identical embeddings and report nDCG@k, Recall@k, MRR@k, precision@1, and latency. Current checked-in run: BEIR SciFact. |
| NoMIRACL Russian retrieval | Russian human-annotated multilingual relevance over compact candidate passages. | implemented | WaveMind / Chroma / Qdrant | Keep same-embedding nDCG@10 at parity, then rerun with sentence-transformers and full MIRACL Russian when disk/service capacity allows it. |
| ANN/VectorDBBench-style local curve | Recall/latency tradeoff for candidate indexes on generated vectors. | implemented | NumPy exact / quantized int8 / Annoy / Qdrant local | Use this as the local engineering curve; official VectorDBBench custom-dataset execution is now runner-ready. |
| Production index profile | Docker-backed 50000-vector profile for persisted FAISS, Qdrant service, and PostgreSQL/pgvector HNSW. | implemented | FAISS / Qdrant service / pgvector | Keep service-mode candidate generation above 0.95 recall@10 and below 10 ms average query latency at 50000 vectors. |
| Production load profile | 100k and 1M service-backed candidate-index checks with p95/p99 latency plus an explicit SLO/cost gate for recall, p99, QPS, replicas, HPA capacity, storage, monthly target cost, and cost per 1M queries. | implemented | Qdrant service / pgvector HNSW / FAISS persisted | Keep 100k Qdrant and 1M persisted FAISS green while tuning Qdrant/pgvector for the same 1M p99 gate. |
| Qdrant 1M HNSW ef sweep | One 1M Qdrant collection queried with multiple hnsw_ef values and the same SLO gate. |
implemented | Qdrant service | Keep the older sweep as a tail-latency regression baseline; the streaming 1M profile now proves the passing path with safe chunks, wait-after-build, and 100 warmup queries. |
| Production streaming load runner | Memory-bounded large-N runner that generates and inserts vectors in batches and measures target-recall, p99, SLO, and cost without storing the full corpus or exact-neighbor matrix in RAM. | implemented | FAISS persisted / FAISS IVF-PQ persisted / Qdrant service streaming / Qdrant sharded service streaming / pgvector streaming | Keep the passing 10M and 50M compressed FAISS profiles green, then complete 10M Qdrant/pgvector service evidence and the real 100M distributed sharded-Qdrant run. |
| Scale readiness profile | Cluster placement, node/zone-loss simulation, quorum report, control-plane majority lease/config revision safety, Kubernetes StatefulSet, HPA, repair CronJob, CRD status subresource, operator readiness/capacity/autoscaling/repair/production-admission/control-plane conditions, service-mode distributed namespace sharding, real HTTP shard transport, sustained mixed HTTP cluster load, replica repair, tombstone-aware delete repair, anti-entropy repair worker, query-vector cache, API batch recall, Redis-compatible shared rate limiting, explicit and batch recall feedback, Memory OS adaptive prewarm/transition-learned predictive prefetch/consolidation/forgetting plus execution-plan safety, replicated runtime, cursor-based active-active namespace delta sync, sustained active-active mesh sync, FastAPI service-region active-active sync, field-only hotness delta sync, field-state CRDT convergence, replicated snapshot/restore with offsite, archive, object-store latest-metadata/download/retention/DR-drill verification, SQLite point-in-time recovery journal replay, query-audit cache prewarm, Redis-compatible shared hot-cache behavior, namespace invalidation, API cache mutation safety on remember/feedback/feedback-batch/forget, structured/multimodal/cross-modal payload retrieval, and temporal event interval/recency retrieval with provenance, persistence, and external precomputed-vector compatibility. | implemented | Mem0 / Zep / LangGraph persistent memory / GraphRAG target adapters | Keep quorum replication, control-plane split-brain rejection, distributed namespace routing, autoscaling manifests, operator status conditions, production admission wiring, scheduled repair, service-mode repair, HTTP shard transport, sustained mixed cluster load, tombstone-aware delete repair, anti-entropy background repair, query-vector cache, API batch recall, shared rate limiting, Memory OS prewarm/transition-learned predictive prefetch/consolidation/forgetting and rollout safety contracts, explicit and batch recall feedback, cursor-bounded namespace-delta sync, sustained active-active mesh convergence, service-region active-active delta endpoints, field-state CRDT merge, repair, local and Redis cache prewarm, mutation-safe shared cache behavior, temporal/cross-modal provenance, precomputed-vector compatibility, offsite/archive/object-store backup lifecycle, SQLite point-in-time recovery, restore drills, and 10M compressed load tests green. |
| Local HTTP cluster smoke | Starts real localhost API-node processes and runs the service-mode sustained mixed workload through HTTP with read_fanout=1, then probes /stats on every node for health/circuit state. |
implemented | WaveMind local API nodes | Keep success, failover, repair, forget, delete suppression, and cluster health at 1.00; GitHub CI uses a 2000 ms p99 ceiling for runner variance while checked-in local evidence stays below 1000 ms. |
| Local HTTP active-active service-region smoke | Starts real localhost API region processes, each serving a replicated WaveMind runtime, then syncs namespace deltas through HTTP export/import endpoints. | implemented | WaveMind local replicated API regions | Keep convergence, delete suppression, and pair-sync success at 1.00, final no-op imports at 0, and p99 below 1500 ms in CI. |
| External HTTP active-active loopback | Starts real localhost API regions and feeds their URLs into the external active-active runner, proving the same URL-based transport contract used by remote deployments. | implemented | WaveMind URL-based API regions | Keep convergence, delete suppression, success, and final no-op sync green in CI while remote Kubernetes/serverless regions are provisioned. |
| External HTTP cluster load runner | benchmarks/http_cluster_load_benchmark.py runs the sustained mixed workload against user-supplied WaveMind API node URLs and reports success rate, failover hit rate, delete suppression, repair count, online query p99, bulk lifecycle batch p99, external /query/batch recall, and slo_pass. |
implemented | WaveMind remote service nodes | Keep the attested pod-DNS artifact green, then repeat the target-specific contract on managed staging endpoints before claiming managed production. |
| External HTTP active-active runner | Manual GitHub workflow and CLI path for user-supplied remote API regions. It writes benchmarks/external_http_active_active_results.json and validates convergence, delete propagation, cursor idempotency, final no-op sync, and p99. |
implemented | WaveMind remote API regions | Check in a real remote Kubernetes/serverless active-active region artifact with convergence, delete suppression, and success at 1.00. |
| Production readiness gate | Machine-readable gate over production artifacts, with pass/action_required/fail criteria. | implemented | WaveMind-only gate | Reach readiness_score 1.000 before claiming complete million-plus production readiness. |
| Memory competitor adapter profile | Generated dynamic-memory scenario wired for external memory frameworks: 50 users, 450 facts, 300 checks, corrections, TTL expiry, namespace collisions, preferences, and token validity. | implemented | Mem0 / Zep / LangGraph persistent memory / GraphRAG static graph | Report real competitor results only when their packages/services are explicitly configured. |
| BEIR | Standard zero-shot information retrieval quality. | planned | Chroma / Qdrant / FAISS | Stay within 0.02 nDCG@10 on identical embeddings. |
| MTEB Retrieval | Separates encoder quality from retrieval-store quality. | planned | Chroma / Qdrant / FAISS | Prove WaveMind does not reduce same-embedding retrieval quality. |
| MIRACL Russian | Multilingual retrieval with Russian relevance judgments. | runner ready | Chroma / Qdrant / FAISS | NoMIRACL Russian compact run is implemented; full-corpus MIRACL Russian remains the next heavier profile. |
| VectorDBBench | Vector database insertion/search/filter/cost-performance benchmark. | runner ready | Chroma / Qdrant / Milvus / Weaviate / Pinecone / FAISS | WaveMind now exports a reproducible custom dataset (train.parquet, test.parquet, neighbors.parquet, scalar_labels.parquet) for official VectorDBBench runs. |
| LoCoMo | Long conversation memory, temporal consistency, multi-hop recall. Retrieval-only runner is implemented for official locomo10.json. |
implemented | Static vector / Chroma / Qdrant | Improve answer generation accuracy on top of the stronger sentence-transformers evidence retrieval run. |
| LongMemEval | Long-term assistant memory with updates and abstention. | implemented retrieval + local Ollama answer smoke | Static vector / Chroma / Qdrant / Mem0-style memory | Add stronger LLM answer quality, abstention, and Chroma/Qdrant RAG answer baselines. |
| LongMemEval-V2 | Web-agent memory: state recall, dynamic state, workflow gotchas. | planned | AgentRunbook-R / Chroma RAG / Qdrant RAG | Prove WaveMind can retrieve compact evidence from agent trajectories. |
| LMEB | Long-horizon memory embedding tasks beyond normal passage retrieval. | planned | Embedding-only baselines / Chroma / Qdrant | Choose the default semantic encoder using memory-specific tasks. |
| RAGBench | Downstream RAG context and answer quality. | planned | Chroma RAG / Qdrant RAG / Pinecone RAG | Show whether stale-memory suppression improves context relevance. |
The planned rows are not claimed wins. They are the public evaluation path WaveMind needs before strong production claims.
Many retrieval benchmarks use the same simple shape:
corpus.jsonl- documents with_id, optionaltitle, andtext.queries.jsonl- queries with_idandtext.qrels/test.tsv- judged relevance rows:query-id,corpus-id,score.
WaveMind includes a BEIR-style runner so the same downloaded dataset can be used for WaveMind, Chroma, and Qdrant:
pip install -e ".[bench]"
python benchmarks/open_retrieval_benchmark.py --dataset ./benchmarks/data/scifact --engines wavemind chroma qdrant --top-k 10That runner reports nDCG@k, Recall@k, MRR@k, precision@1, average
latency, and p95 latency. It intentionally uses the same WaveMind encoder for
all engines, so the comparison is about retrieval/index behavior rather than
which embedding model each project chooses by default.
Checked-in BEIR SciFact result:
5183 documents, 300 test queries, HashingTextEncoder, top-k 10.
This is a public retrieval sanity check, not the main agent-memory proof.
Full machine-readable result: benchmarks/open_retrieval_scifact_results.json.
| engine | nDCG@10 | Recall@10 | MRR@10 | precision@1 | avg latency | p95 latency |
|---|---|---|---|---|---|---|
| WaveMind | 0.354 | 0.482 | 0.317 | 0.240 | 117.02 ms | 256.57 ms |
| Chroma | 0.350 | 0.467 | 0.315 | 0.243 | 1.79 ms | 2.39 ms |
| Qdrant | 0.354 | 0.482 | 0.317 | 0.240 | 17.71 ms | 23.28 ms |
Read this result narrowly: WaveMind preserves same-embedding retrieval quality on a real public dataset, but its current exact path is far slower than Chroma. Qdrant local preserves the same ranking quality and is much faster than the WaveMind NumPy exact path. The engineering target is a FAISS/Annoy candidate index with WaveMind's dynamic field policy applied only as a top-k re-ranker.
WaveMind includes a compact multilingual retrieval runner for
NoMIRACL, the negative-aware
MIRACL relevance dataset. The checked-in run uses Russian test.relevant
queries and the compact Russian candidate corpus. It is not a full-corpus
MIRACL run; it is a reproducible multilingual relevance benchmark small enough
to run on a local machine.
python benchmarks/nomiracl_russian_benchmark.py --download --dataset benchmarks/data/nomiracl-russian --engines wavemind chroma qdrant --top-k 10 --limit-queries 200 --limit-corpus 5000 --output benchmarks/nomiracl_russian_results.jsonChecked-in NoMIRACL Russian result:
200 Russian queries, 5000 compact candidate passages,
HashingTextEncoder, top-k 10. Full machine-readable result:
benchmarks/nomiracl_russian_results.json.
| engine | nDCG@10 | Recall@10 | MRR@10 | precision@1 | avg latency | p95 latency |
|---|---|---|---|---|---|---|
| WaveMind | 0.434 | 0.516 | 0.489 | 0.410 | 10.22 ms | 15.53 ms |
| Chroma | 0.435 | 0.519 | 0.490 | 0.410 | 2.60 ms | 3.44 ms |
| Qdrant | 0.434 | 0.516 | 0.489 | 0.410 | 18.86 ms | 24.08 ms |
Read this as multilingual same-embedding parity, not as a claim that the hash
encoder is the best Russian semantic model. The next stronger run should use
sentence-transformers on the same NoMIRACL split, then full MIRACL Russian
when there is enough disk/service capacity.
WaveMind now includes a retrieval-only runner for the public
LoCoMo dataset. It treats LoCoMo
conversation turns as memories and LoCoMo QA evidence dialog IDs as relevance
labels. This measures the memory layer before any LLM answer-generation noise.
Run it on the official locomo10.json file:
mkdir -p benchmarks/data
curl -L https://raw.githubusercontent.com/snap-research/locomo/main/data/locomo10.json -o benchmarks/data/locomo10.json
python benchmarks/locomo_memory_benchmark.py --dataset benchmarks/data/locomo10.json --engines wavemind static chroma qdrant --top-k 5 --output benchmarks/locomo_evidence_results.jsonMetrics reported:
evidence_recall@k- whether the labeled LoCoMo evidence turns appear in the returned memory block.precision@1- whether the first returned memory is labeled evidence.MRR@k- how high the first relevant evidence turn appears.context_budget_saved- how much smaller the returned evidence block is than the full conversation memory.avg_latency_msandp95_latency_ms- retrieval latency only.
If Chroma or Qdrant are not installed, use the baseline-only command:
python benchmarks/locomo_memory_benchmark.py --dataset benchmarks/data/locomo10.json --engines wavemind static --top-k 5For multi-tenant local deployments, ShardedWaveMind routes namespaces across
multiple SQLite files:
from wavemind import ShardedWaveMind
memory = ShardedWaveMind(root_path="./state/wavemind-shards", shard_count=16)
memory.remember("Tenant A prefers short support replies.", namespace="tenant:a")
memory.remember("Tenant B tracks trading research.", namespace="tenant:b")
print(memory.query("support replies", namespace="tenant:a", top_k=3))
print(memory.stats())
memory.close()For HA-style local or service-mode deployments, ReplicatedWaveMind writes each
namespace to a deterministic replica set and enforces read/write quorum:
from wavemind import ReplicatedSnapshotWorker, ReplicatedWaveMind
memory = ReplicatedWaveMind(
root_path="./state/wavemind-replicas",
nodes=[
{"id": "node-a", "address": "10.0.0.1:8000", "zone": "zone-a"},
{"id": "node-b", "address": "10.0.0.2:8000", "zone": "zone-b"},
{"id": "node-c", "address": "10.0.0.3:8000", "zone": "zone-c"},
],
replication_factor=3,
)
memory.remember("Tenant A prefers short support replies.", namespace="tenant:a")
print(memory.query("support replies", namespace="tenant:a", top_k=3))
memory.set_node_available("node-a", False)
print(memory.query("support replies", namespace="tenant:a", top_k=3))
memory.close()The runtime uses separate durable stores per node, stable replica keys, operation
metadata, quorum writes, quorum reads, merged replica results, tombstone-aware
delete propagation, and repair_namespace() for recovered replicas. It is the
production foundation for namespace-level HA and eventual-consistency behavior;
for full consensus across independent network services, deploy WaveMind with
Postgres/Qdrant/ops-layer replication.
For multi-region active-active experiments, use cursor-based namespace delta sync. The first call can transfer the full namespace; later calls use the returned cursor and transfer only new records, tombstones, or field-state keys:
from wavemind import sync_namespace_delta
region_a.remember("Tenant A billing preference.", namespace="tenant:a")
first = sync_namespace_delta(region_a, region_b, "tenant:a")
region_a.remember("Tenant A latency preference.", namespace="tenant:a")
next_page = sync_namespace_delta(
region_a,
region_b,
"tenant:a",
since=first.to_cursor,
)
region_a.forget(text="Tenant A billing preference.", namespace="tenant:a")
delete_sync = sync_namespace_delta(
region_a,
region_b,
"tenant:a",
since=next_page.to_cursor,
)The delta contains active records plus tombstones. Import is idempotent and tombstone-aware, so a stale region export cannot resurrect a deleted memory. The replicated runtime also carries field-state CRDT deltas for active-active hotness/suppression signals. Each delta includes per-actor watermarks, so a region can audit which writers are covered by the field state it has received, and detect missing actors or lagging field-state replicas before serving stale priority. Recall-only hotness changes can sync without resending records:
region_a.query("latency preference", namespace="tenant:a")
field_only = sync_namespace_delta(
region_a,
region_b,
"tenant:a",
since=delete_sync.to_cursor,
)This lets regions converge on dynamic memory priority without double-counting the same signal when a delta is replayed.
For more than two regions, ActiveActiveSyncWorker keeps per-pair cursors and
runs bounded mesh sync cycles across namespaces:
from wavemind import ActiveActiveSyncWorker
worker = ActiveActiveSyncWorker(
{
"us-east": region_a,
"eu-west": region_b,
"ap-south": region_c,
}
)
report = worker.run_once(["tenant:a", "tenant:b"], bidirectional=True)
assert report.okThe checked-in scale-readiness profile exercises 3 independent regions, 3
namespaces, 18 writes, 90 region-pair syncs, tombstone propagation, field-only
hotness sync, and a final no-op sync that imports 0 records. This is still a
local independent-region profile, not a claim that remote production regions
have already been load-tested.
For operational recovery, ReplicatedSnapshotWorker creates a checksummed
replicated snapshot, verifies an optional offsite mirror, writes a verified
.tar.gz archive, can upload that archive to S3-compatible storage, and
ReplicatedObjectStoreDrillWorker can run a recovery drill from the newest or
exact remote archive:
job = ReplicatedSnapshotWorker(memory).run_once(
destination="./backups/replicated",
offsite_destination="./offsite/replicated",
archive_destination="./archives/replicated",
object_store_destination="s3://my-bucket/wavemind/prod",
keep_last=7,
)
restored, report = ReplicatedWaveMind.restore_snapshot_archive(
job.archive_path,
"./state/restored-replicas",
)The checked-in scale-readiness profile verifies manifest checksums, verifies the offsite mirror, verifies the portable archive, verifies S3-compatible object upload metadata, downloads the latest remote archive, verifies its SHA-256 against object metadata, restores three replica files from that archive, then disables the restored primary and confirms the memory is still recalled from the remaining replicas.
Checked-in official LoCoMo retrieval result:
10 conversations, 5882 memory turns, 1977 evidence-labeled questions,
HashingTextEncoder, top-k 5. Full machine-readable result:
benchmarks/locomo_evidence_results.json.
| engine | evidence recall@5 | precision@1 | MRR@5 | avg latency | p95 latency |
|---|---|---|---|---|---|
| WaveMind | 0.386 | 0.239 | 0.307 | 3.95 ms | 7.44 ms |
| Static vector | 0.263 | 0.133 | 0.189 | 1.94 ms | 3.87 ms |
| Chroma static | 0.257 | 0.129 | 0.185 | 7.03 ms | 9.74 ms |
| Qdrant static | 0.263 | 0.133 | 0.189 | 147.58 ms | 210.23 ms |
Checked-in semantic LoCoMo run:
Same official data, same engines, but with
sentence-transformers/paraphrase-multilingual-mpnet-base-v2. Full
machine-readable result: benchmarks/locomo_sentence_evidence_results.json.
| engine | evidence recall@5 | precision@1 | MRR@5 | avg latency | p95 latency |
|---|---|---|---|---|---|
| WaveMind | 0.547 | 0.333 | 0.432 | 3.44 ms | 5.56 ms |
| Static vector | 0.409 | 0.219 | 0.305 | 1.25 ms | 2.05 ms |
| Chroma static | 0.407 | 0.218 | 0.304 | 4.97 ms | 6.30 ms |
| Qdrant static | 0.409 | 0.219 | 0.305 | 124.34 ms | 149.72 ms |
Read this as retrieval-only evidence quality, not full QA quality. It uses the same embeddings for every engine inside each table. The sentence-transformers run is the stronger evidence-quality number: WaveMind improves recall over static vector-store retrieval, while Chroma remains the fastest retrieval backend. The next LoCoMo step is answer generation and faithfulness with a local LLM on top of retrieved evidence.
WaveMind also includes a retrieval-only runner for the official LongMemEval format. It indexes each question's long chat history and measures whether the expected evidence sessions or turns are retrieved before answer generation.
Run the full session-level retrieval benchmark:
python benchmarks/longmemeval_memory_benchmark.py --dataset benchmarks/data/longmemeval_s_cleaned.json --engines wavemind static chroma qdrant --granularity session --top-k 5 --output benchmarks/longmemeval_evidence_results.jsonChecked-in official LongMemEval-S retrieval result:
470 non-abstention questions from longmemeval_s_cleaned.json,
22419 session memories, HashingTextEncoder, top-k 5. Full machine-readable
result: benchmarks/longmemeval_evidence_results.json.
| engine | evidence recall@5 | precision@1 | MRR@5 | context saved | avg latency | p95 latency |
|---|---|---|---|---|---|---|
| WaveMind | 0.782 | 0.696 | 0.762 | 0.869 | 7.27 ms | 9.14 ms |
| Static vector | 0.520 | 0.355 | 0.464 | 0.890 | 0.08 ms | 0.10 ms |
| Chroma static | 0.518 | 0.355 | 0.464 | 0.890 | 15.96 ms | 18.68 ms |
| Qdrant static | 0.520 | 0.355 | 0.464 | 0.890 | 398.48 ms | 432.88 ms |
The Chroma and Qdrant baselines now use the same namespace/payload scope as WaveMind. Qdrant is run in local embedded mode; the Qdrant client warns that local mode is not recommended above 20000 points, so this latency should not be read as a service-mode Qdrant result.
Answer-generation runner with local Ollama:
python benchmarks/longmemeval_answer_benchmark.py --dataset benchmarks/data/longmemeval_s_cleaned.json --provider ollama --model YOUR_LOCAL_MODEL --engines wavemind chroma qdrant --top-k 5 --output benchmarks/longmemeval_answer_results.jsonChecked-in local answer-generation smoke runs:
50 non-abstention LongMemEval-S questions, compact retrieved evidence,
same HashingTextEncoder, same local Ollama model, top-k 5. Full machine-readable results:
benchmarks/longmemeval_answer_qwen25_0_5b_50_results.json and
benchmarks/longmemeval_answer_qwen25_1_5b_50_results.json.
| system | questions | evidence recall@5 | exact match | contains answer | token F1 | avg retrieval | avg generation |
|---|---|---|---|---|---|---|---|
WaveMind + Ollama qwen2.5:0.5b |
50 | 0.920 | 0.120 | 0.180 | 0.183 | 2.98 ms | 1428.20 ms |
Chroma static + Ollama qwen2.5:0.5b |
50 | 0.600 | 0.100 | 0.120 | 0.126 | 4.10 ms | 1234.69 ms |
Qdrant static + Ollama qwen2.5:0.5b |
50 | 0.600 | 0.100 | 0.120 | 0.126 | 63.80 ms | 893.48 ms |
WaveMind + Ollama qwen2.5:1.5b |
50 | 0.920 | 0.240 | 0.380 | 0.333 | 2.00 ms | 2153.00 ms |
Chroma static + Ollama qwen2.5:1.5b |
50 | 0.600 | 0.120 | 0.160 | 0.170 | 7.05 ms | 2082.38 ms |
Qdrant static + Ollama qwen2.5:1.5b |
50 | 0.600 | 0.120 | 0.160 | 0.170 | 100.20 ms | 758.11 ms |
There is also an extractive smoke run that does not require a model:
benchmarks/longmemeval_answer_extractive_20_results.json. It is only a runner
check, not a meaningful final answer-quality benchmark. The Ollama runs are real
local LLM runs, but still lightweight smoke results rather than official
LongMemEval leaderboard scores.
WaveMind includes a local ANN/VectorDBBench-style curve for candidate indexes.
It generates normalized vectors, queries with noisy copies, and measures
recall@10 against exact cosine neighbors.
python benchmarks/ann_index_curve_benchmark.py --sizes 1000 5000 10000 50000 --dim 128 --queries 100 --top-k 10 --engines numpy quantized annoy faiss qdrant-local --output benchmarks/ann_index_curve_results.jsonAdd pgvector to --engines when WAVEMIND_PGVECTOR_DSN points at a
PostgreSQL database with pgvector enabled. Use pgvector-exact for recall
audits and pgvector-iterative for the HNSW + filtered-collection tuning path
on pgvector builds that support iterative scans.
Add qdrant-service when WAVEMIND_QDRANT_URL points at a running Qdrant
service. Add faiss-persisted when WAVEMIND_FAISS_PATH points at the FAISS
snapshot file to validate persisted-index startup behavior.
VectorDBBench custom dataset export:
python benchmarks/vectordbbench_dataset.py --vectors 10000 --queries 100 --dim 128 --top-k 10 --output-dir state/vectordbbench-wavemind --manifest benchmarks/vectordbbench_dataset_manifest.jsonThis writes the standard custom-dataset files used by VectorDBBench:
train.parquet, test.parquet, neighbors.parquet, and
scalar_labels.parquet; the ground-truth file uses the neighbors_id column.
The parquet files are generated under state/ so the
repository does not carry bulky binary benchmark data; the checked-in manifest
records the exact parameters and reproduction command.
Reproducible Docker production profile:
docker compose -f examples/production-index-profile/docker-compose.yml up -d qdrant postgres
docker compose -f examples/production-index-profile/docker-compose.yml run --rm benchmark
docker compose -f examples/production-index-profile/docker-compose.yml downChecked-in 50000-vector point:
| engine | recall@10 | avg latency | p95 latency | build |
|---|---|---|---|---|
| WaveMind numpy | 1.000 | 1.99 ms | 1.45 ms | 395.1 ms |
| WaveMind quantized | 0.934 | 16.27 ms | 20.15 ms | 1019.3 ms |
| WaveMind annoy | 0.730 | 3.21 ms | 4.17 ms | 2300.3 ms |
| WaveMind faiss | 1.000 | 80.47 ms | 2.42 ms | 458.3 ms |
| Qdrant local | 1.000 | 33.82 ms | 39.22 ms | 8336.8 ms |
Checked-in production 50000-vector point:
| engine | recall@10 | avg latency | p95 latency | build |
|---|---|---|---|---|
| WaveMind faiss-persisted | 1.000 | 3.52 ms | 7.88 ms | 715.9 ms |
| Qdrant service | 1.000 | 4.41 ms | 5.93 ms | 12269.8 ms |
| WaveMind pgvector | 0.811 | 10.95 ms | 15.69 ms | 185048.9 ms |
Checked-in pgvector tuning 50000-vector point:
| engine | recall@10 | avg latency | p95 latency | p99 latency | build |
|---|---|---|---|---|---|
| Qdrant service | 1.000 | 9.14 ms | 17.31 ms | 17.84 ms | 14786.2 ms |
| WaveMind pgvector | 0.834 | 14.52 ms | 19.59 ms | 21.08 ms | 200277.7 ms |
| WaveMind pgvector-exact | 1.000 | 55.72 ms | 66.91 ms | 76.98 ms | 211611.0 ms |
| WaveMind pgvector-iterative | 0.970 | 38.32 ms | 48.02 ms | 55.19 ms | 189784.3 ms |
Checked-in production load points:
python benchmarks/production_load_benchmark.py --sizes 100000 --dim 128 --queries 100 --top-k 10 --engines qdrant-service pgvector pgvector-exact pgvector-iterative faiss-persisted
python benchmarks/production_load_benchmark.py --sizes 1000000 --dim 128 --queries 100 --top-k 10 --engines qdrant-service --output benchmarks/production_load_qdrant_1m_tuned_results.json
python benchmarks/production_load_benchmark.py --sizes 1000000 --dim 128 --queries 100 --top-k 10 --engines faiss-persisted --output benchmarks/production_load_faiss_1m_results.json| vectors | engine | recall@10 | avg latency | p95 latency | p99 latency | SLO | required replicas | autoscaled capacity | build |
|---|---|---|---|---|---|---|---|---|---|
| 100000 | Qdrant service | 1.000 | 10.28 ms | 18.97 ms | 21.26 ms | pass | 2 | 1635.0 qps | 27439.3 ms |
| 100000 | WaveMind pgvector | 0.736 | 17.76 ms | 23.48 ms | - | fail: recall | 3 | 945.9 qps | 455703.7 ms |
| 100000 | WaveMind faiss-persisted | skipped | - | - | - | skipped | - | - | - |
| 1000000 | WaveMind faiss-persisted | 1.000 | 39.12 ms | 45.29 ms | 57.71 ms | scale required | 6 | 429.5 qps | 20788.1 ms |
| 1000000 | Qdrant service tuned | 0.984 | 82.57 ms | 125.99 ms | 137.86 ms | fail: p99 | 12 | 203.5 qps | 441775.0 ms |
| 1000000 | Qdrant hnsw_ef=2048 sweep point |
0.977 | 64.76 ms | 91.18 ms | 103.77 ms | fail: p99 | 10 | 259.4 qps | 451912.4 ms |
Read this as an engineering curve, not an official VectorDBBench result. NumPy
exact improved after switching top-k selection away from full sorting. Annoy is
fast at 50000 vectors but loses too much recall with the current settings. The
quantized backend now stores the candidate matrix as int8 and uses an
int32-safe accumulator; it keeps 0.934 recall@10 on this run, but it is still a
memory-footprint baseline rather than a latency win. FAISS flat preserves exact
recall and low warm p95, but the local curve includes a cold-start average
outlier, so production claims should use the persisted/service profiles below.
FAISS persistence and service-mode Qdrant now both preserve exact recall at
50000 generated vectors. The checked-in pgvector/HNSW profile uses
WAVEMIND_PGVECTOR_EF_SEARCH=400, which improves recall materially but still
misses the 0.95 production target and is slower than the other two profiles.
The runner now has explicit pgvector-exact and pgvector-iterative engines:
use exact mode as the correctness floor, then use iterative HNSW scan to tune
the production latency/recall tradeoff without hiding approximation loss.
The 100k load profile shows Qdrant service is already viable for candidate
generation on the tested machine under the checked-in SLO gate. The 1M
persisted-FAISS profile passes recall and p99 with a 100-query run. The older
1M Qdrant load profile still records a p99 tail failure, but the newer
streaming 1M Qdrant run passes the same p99 target after safe upsert chunking,
wait-after-build, and 100 warmup queries.
If a required package, service, or environment variable is missing, the runner
marks that engine as skipped instead of silently falling back to another
backend.
WaveMind includes a generated dynamic-memory adapter profile for Mem0, Zep,
LangGraph persistent memory, and a local GraphRAG-style static graph baseline.
The checked-in run covers 50 synthetic users, 450 facts, and 300 checks across
corrections, TTL expiry, namespace collisions, preference recall, and
token-validity questions. Mem0, LangGraph, and GraphRAG-style run in the
checked-in profile; Zep is marked skipped until a dedicated service/API key is
configured.
python benchmarks/memory_competitor_benchmark.py --engines wavemind mem0 zep langgraph graphrag --generated-users 50 --namespaces 10For a live Zep run, install zep-cloud for the current Zep Cloud Graph API
or zep-python for a legacy/OSS-compatible Zep service, then set either
ZEP_API_URL or ZEP_API_KEY. The benchmark creates temporary graphs/sessions
and deletes them after the run.
| engine | precision@1 | precision@3 | stale suppression | avg latency |
|---|---|---|---|---|
| WaveMind | 0.80 | 0.94 | 1.00 | 14.55 ms |
| Mem0 | 0.68 | 0.99 | 0.60 | 39.32 ms |
| Zep | skipped | - | - | - |
| LangGraph persistent memory | 0.80 | 0.95 | 1.00 | 1.13 ms |
| GraphRAG static graph | 0.85 | 0.96 | 1.00 | 0.08 ms |
Field memory dynamics benchmark:
13 memories, 5 conflicting-fact queries, deterministic local encoder.
This benchmark isolates the MemoryFieldGraph: related memories can spread
activation, newer conflicting memories inhibit stale facts, graph energy decays,
and active clusters can surface and persist concept memories.
Full machine-readable result: benchmarks/field_memory_dynamics_results.json.
| engine | precision@1 | precision@3 | stale suppression | concept formation | concept consolidation | decay ratio | avg latency |
|---|---|---|---|---|---|---|---|
| WaveMind graph | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 0.81 | 0.82 ms |
| WaveMind static | 0.20 | 1.00 | 0.20 | 0.00 | 0.00 | 0.00 | 0.48 ms |
Run locally from a cloned repository:
python benchmarks/field_memory_dynamics_benchmark.pyLong-term memory evidence benchmark:
200 memories, 8 evidence queries, same HashingTextEncoder embeddings.
This benchmark asks a stricter agent-memory question than static retrieval:
did the system return the right evidence while suppressing stale, corrected,
expired, or cross-user evidence?
Full machine-readable result: benchmarks/long_memory_evidence_results.json.
| engine | evidence recall@5 | precision@1 | stale suppression | context saved | avg latency |
|---|---|---|---|---|---|
| WaveMind | 1.00 | 1.00 | 1.00 | 0.87 | 6.10 ms |
| Static vector | 1.00 | 0.57 | 0.00 | 0.88 | 0.65 ms |
Run locally from a cloned repository:
python benchmarks/long_memory_evidence_benchmark.py --dataset synthetic --engines wavemind static --memories 200 --top-k 5To compare the same normalized benchmark with Chroma or Qdrant, install the benchmark extras and include those engines:
pip install -e ".[bench]"
python benchmarks/long_memory_evidence_benchmark.py --dataset synthetic --engines wavemind chroma qdrant --memories 200 --top-k 5Real Russian sentences from Tatoeba, 50 one-word queries, NumPy exact index.
| metric | hash | sentence-transformers |
|---|---|---|
| precision@1 | 1.00 | 1.00 |
| precision@3 | 1.00 | 1.00 |
| avg query | 0.49 ms | 52.84 ms |
Capacity check with the hash encoder:
| memories | precision@1 | precision@3 | avg query |
|---|---|---|---|
| 200 | 1.00 | 1.00 | 0.49 ms |
| 1000 | 0.88 | 1.00 | 1.50 ms |
| 5000 | 0.72 | 0.88 | 5.68 ms |
Run locally from a cloned repository:
python benchmarks/ru_sentences_benchmark.py --sentences 200 --queries 50 --encoder hash --index numpy
python benchmarks/ru_sentences_benchmark.py --sentences 200 --queries 50 --encoder sentence --index numpyAgent-memory benchmark against Chroma:
200 Russian user facts, 50 natural-language questions, same precomputed HashingTextEncoder embeddings for WaveMind and Chroma.
Full machine-readable result: benchmarks/agent_memory_results.json.
This is a static retrieval benchmark. It measures baseline ranking and latency, not hotness, TTL, repeated recall, or memory aging.
| engine | precision@1 | precision@3 | avg latency |
|---|---|---|---|
| WaveMind | 0.82 | 0.90 | 2.25 ms |
| Chroma | 0.82 | 0.88 | 0.93 ms |
WaveMind-only capacity checks from the current ranking path:
| scenario | memories | precision@1 | precision@3 | avg latency | p95 latency |
|---|---|---|---|---|---|
| static agent facts | 200 | 0.96 | 0.98 | 4.05 ms | 8.18 ms |
| static agent facts | 1000 | 0.96 | 0.98 | 3.53 ms | 5.20 ms |
| static agent facts | 5000 | 0.94 | 0.98 | 13.71 ms | 17.20 ms |
| dynamic memory policy | 200 | 1.00 | 1.00 | 38.40 ms | 41.14 ms |
| dynamic memory policy | 1000 | 1.00 | 1.00 | 54.29 ms | 72.38 ms |
| dynamic memory policy | 5000 | 1.00 | 1.00 | 48.36 ms | 86.13 ms |
Machine-readable local capacity result: benchmarks/wavemind_capacity_results.json.
These capacity checks are WaveMind-only because the local restricted environment did not have Chroma installed.
Run locally from a cloned repository:
pip install -e ".[bench]"
python benchmarks/agent_memory_benchmark.py --engines wavemind chroma --facts 200 --queries 50Dynamic agent-memory benchmark:
200 memories, 8 checks, same precomputed HashingTextEncoder embeddings.
This benchmark exercises hot memory, TTL, corrections, and namespace isolation.
WaveMind applies its built-in memory policy. Chroma static is a plain vector-store baseline without application-layer TTL, delete handling, namespace filters, or recall reinforcement.
Full machine-readable result: benchmarks/dynamic_memory_results.json.
| engine | precision@1 | precision@3 | stale suppression | avg latency |
|---|---|---|---|---|
| WaveMind | 1.00 | 1.00 | 1.00 | 25.26 ms |
| Chroma static | 0.57 | 1.00 | 0.00 | 1.75 ms |
Category success:
| behavior | WaveMind | Chroma static |
|---|---|---|
| hot memory | 1.00 | 0.50 |
| TTL | 1.00 | 0.00 |
| correction | 1.00 | 0.00 |
| namespace isolation | 1.00 | 0.00 |
Run locally from a cloned repository:
pip install -e ".[bench]"
python benchmarks/dynamic_memory_benchmark.py --engines wavemind chroma --memories 200| feature | WaveMind | Chroma | Qdrant |
|---|---|---|---|
| Primary role | Dynamic memory engine | Embedding database | Production vector database |
| Local SQLite persistence | Yes | Yes | No, separate service/storage |
| HTTP API | FastAPI included | Included | Included |
| Audit log / metrics | SQLite audit events plus /metrics |
App-layer only | App-layer / service metrics |
| Dynamic memory priority | Wave-field hotness, TTL, priority | Metadata/filter driven | Payload/filter driven |
| Built-in forgetting | TTL and explicit forget | Manual delete/filtering | Manual delete/filtering |
| Best fit | Small to medium memory streams with dynamic recall | Local RAG apps and prototypes | Large-scale vector search |
| Scale target today | Local exact mode for small streams; FAISS/Qdrant/pgvector plus replicated namespaces for production paths | Larger than WaveMind local exact mode | Production vector scale |
WaveMind is not trying to replace dedicated vector databases at scale. The intended product gap is dynamic priority: frequently used memories can become hotter while old or low-priority memories fade. For static RAG over large document collections, use a mature vector database. For memory that needs persistence, scoped recall, TTL, forgetting, and reinforcement, WaveMind is designed to sit above or beside the vector index.
If you already use Chroma for local memory, see the practical migration guide:
docs/CHROMA_MIGRATION.md.
- Optimal capacity on the current NumPy exact index is up to 1000 records.
- At 5000 records, one-word
precision@1is currently 0.72 with the hash encoder; many misses are ambiguous queries where another sentence containing the same word ranks first. - For
N > 5000, the NumPy exact index is still reliable but scales linearly. Annoy is faster at 50000 vectors in the local curve, but current recall is only0.730; thequantizedbackend reaches0.934recall@10 with int8 storage and int32-safe scoring, but is still slower than NumPy on this workload. Use FAISS or a production vector service before claiming large-scale ANN quality. - Run
wavemind scale-plan --target-memories <N>before growing a deployment. It is a guardrail, not a benchmark replacement: it tells you when NumPy is no longer the right candidate index and which checks to run next. sentence-transformers/paraphrase-multilingual-mpnet-base-v2requires about 420 MB of model files. Benchmark runners cache embeddings so retrieval latency is measured separately from model encoding latency.- The Chroma comparison currently uses shared precomputed hash embeddings to isolate retrieval/ranking behavior; semantic model comparisons should be run separately.
- The BEIR SciFact run uses the hash encoder to isolate index/retrieval behavior. It is not a semantic embedding leaderboard result.
- On BEIR SciFact, WaveMind and Qdrant match on hash-encoder
nDCG@10, while Chroma is much faster. The next index milestone is FAISS/Annoy candidate generation plus WaveMind top-k re-ranking. - The LoCoMo results are retrieval-only evidence results, not final answer-quality scores. The sentence-transformers run is stronger than the hash run, but still needs answer generation and faithfulness checks.
- In the 200-fact agent benchmark, Chroma is faster on average while WaveMind is slightly higher at
precision@3. - The dynamic benchmark currently compares WaveMind against a static Chroma baseline. Chroma and Qdrant can implement similar behavior with extra application-layer metadata policy, deletes, filters, and reinforcement logic.
MemoryFieldGraphis a discrete graph over stored memories, not a continuous mathematical field. Its current build path should be optimized with incremental edge updates before large production use.- Kubernetes operator reconciliation now has durable Lease/etcd-backed leader
election with CAS and failover between operator replicas. The separate
ControlPlaneConsensusprofile remains a deterministic config-safety guard, not an implementation of a replicated Raft log for the WaveMind data plane. - The checked Kubernetes network drill writes
256deterministic memories through four pod-DNS API nodes in three worker zones, physically pauses one kind worker, detects the unreachable replica, preserves1.00quorum recall, and returns to1.00recall with no failed nodes after recovery. This is real non-loopback CI evidence, but it is still ephemeral kind evidence rather than remote multi-region production admission. Seebenchmarks/kubernetes_cluster_network_smoke_results.jsonand its traceable GitHub Actions run. - The checked active-active Kubernetes drill uses three separate replicated
region APIs with persistent volumes in three worker zones. During a physical
zone-B worker pause, regions A and C continue writes and delete propagation at
1.00convergence; the recovered region converges without resurrecting the deleted memory and the final sync is a no-op. This remains ephemeral kind evidence rather than independent remote-region admission. Seebenchmarks/kubernetes_active_active_region_smoke_results.jsonand its GitHub Actions run. - pgvector is a candidate-index backend. PostgreSQL source-of-truth storage is
also available separately. A Postgres-native PITR runbook/preflight now exists
through
wavemind postgres-pitr-planandbenchmarks/postgres_pitr_plan.json; migrations, real managed-Postgres PITR drill evidence, and larger service benchmark profiles still need more real deployment coverage. - The Qdrant backend is also a candidate-index backend. WaveMind rebuilds it from SQLite on load/build, so large service-mode deployments still need a measured rebuild strategy and index-health monitoring.
- The persisted FAISS backend validates a snapshot against current memory ids
and avoids unnecessary FAISS rebuilds when the snapshot matches. FAISS itself
is a single-node flat-index path; use
ReplicatedWaveMindor external database/service replication when that is not enough. - The current cross-modal layer supports deterministic descriptor embeddings, a strict precomputed-vector path for externally computed CLIP/audio/video/3D embeddings, and an optional sentence-transformers backend for CLIP-style local image/text retrieval. Audio, video, and 3D perception still require external embeddings or strong descriptors until dedicated backends are benchmarked.
wavemind multimodal-admissionkeeps production multimodal claims locked until an external encoder/object-store benchmark artifact proves real image/audio/video/3D quality, cross-modal routing, object-store verification, persistence, provenance, p99 query latency, encode p95, and error-rate thresholds. Usewavemind multimodal-external-evidenceto generate that artifact from a real external manifest; no fixture unlocks this claim.- The
quantizedbackend is an explicit int8 candidate-index experiment. It reduces vector precision, stores the local candidate matrix compactly, uses an int32 accumulator to avoid dot-product overflow, and must be benchmarked per workload before use. - The synthetic long-term memory evidence benchmark is useful for regression and product-shape proof, but public claims should lean on LoCoMo and LongMemEval instead.
- The main LongMemEval evidence result is retrieval-only. The checked-in Ollama answer-generation comparison now includes WaveMind, Chroma static, and Qdrant static over 50 questions, but it is still not a full LongMemEval leaderboard-equivalent score.
- Qdrant baselines in this README use embedded local mode. Qdrant itself warns that local mode is not recommended above 20000 points; use the
qdrant-servicebenchmark profile before making production latency claims. - The tuned 1M Qdrant streaming result depends on safe upsert chunking,
30seconds wait-after-build, and100warmup queries. The cold 1M run misses the p99 SLO, so production Qdrant claims must specify warmup/tuning behavior. - The Qdrant streaming path has a real service smoke, a real two-service sharded smoke, a tuned 1M passing run, and checked single-service plus sharded 10M preflight contracts. It is not a completed 10M Qdrant benchmark until
benchmarks/production_streaming_load_qdrant_10m_results.jsonandbenchmarks/production_streaming_load_qdrant_sharded_10m_results.jsonare produced by real runs. - The pgvector streaming path has a real service smoke and a checked 10M preflight contract. It is not a completed 10M pgvector benchmark until
benchmarks/production_streaming_load_pgvector_10m_results.jsonis produced by a real run. - The production cost model is an engineering estimate from checked-in benchmark parameters: required replicas, target QPS, replica hourly cost, vector storage, and payload storage. It is not a cloud-provider bill and must be recalibrated for real hardware.
- MTEB, MIRACL, LMEB, official VectorDBBench, and RAGBench are listed as the public benchmark roadmap, not as completed results yet.
- Local Ollama answer generation now works with
qwen2.5:0.5bandqwen2.5:1.5b; WaveMind leads the checked-in Chroma/Qdrant smoke comparison, but answer quality is still limited by small-model reasoning and should be rerun with stronger local/API models before making product claims. - Public benchmark adapters require optional datasets, heavier dependencies, or running services. They are intentionally outside the minimal
pip install wavemindpath. - Dynamic memory is slower than static Chroma in the current local benchmark: 25.26 ms vs 1.75 ms average query latency on this machine.
- Current WaveMind-only dynamic checks keep
precision@1at 1.00 through 5000 memories, but average latency is around 48-54 ms. The next optimization target is field/re-ranking latency, not basic recall quality.
Full roadmap: docs/ROADMAP.md.
Launch and positioning kit: docs/LAUNCH_KIT.md.
Near-term priorities:
- Service-mode Qdrant, pgvector, and persisted-FAISS benchmark runs on a real production-like machine, with SLO and cost gates checked into the repo.
- Migration tooling and operational docs for Postgres source-of-truth storage.
- Tune the quantized int8 backend so compression does not cost more latency than exact NumPy on common workloads.
- Service-mode Qdrant and FAISS latency baselines using the explicit Qdrant backend, not only the standalone Qdrant benchmark baseline.
- LoCoMo and LongMemEval answer-quality evaluation, not retrieval only.
- Harden framework adapters: LangGraph, LlamaIndex, CrewAI, AutoGen, OpenClaw, and HTTP-only sidecar use.
- Faster dynamic re-ranking through smaller candidate windows, caching, and background updates.
- Better production operations: OpenTelemetry, SQLite point-in-time recovery, and replicated offsite snapshot jobs with verified portable archives, S3-compatible upload, latest-archive lookup, restore from latest, object-store DR drill, object-store retention, and a Postgres PITR runbook/preflight are implemented; richer latency histograms, index-health metrics, alerting examples, real cloud disaster-recovery drills, and a real managed-Postgres PITR drill report are next.
Longer-term direction:
- scale from thousands of memories to 100k-1M on one node;
- keep SQLite as the local source of truth while adding Postgres and external vector backends for production;
- evolve
MemoryFieldGraphfrom a regression-tested graph into a stronger field-memory model with excitation, inhibition, decay, and consolidation; - expand the built-in multimodal backend beyond CLIP-style local image/text retrieval into benchmarked audio/video/3D encoders while keeping the same provenance-preserving payload API;
- build enterprise features only after benchmarked retrieval, latency, and answer-quality evidence are solid.
Contributing guide: CONTRIBUTING.md.
Useful contribution paths:
- add reproducible benchmark adapters and checked-in result JSON;
- improve FAISS, Qdrant, pgvector, or other candidate-index backends;
- add examples for LangGraph, LlamaIndex, CrewAI, AutoGen, OpenClaw, and HTTP-only sidecar deployments;
- improve dynamic memory behavior around TTL, corrections, namespaces, graph excitation/inhibition, and consolidation;
- harden production operations: backups, audit logs, metrics, tracing, and migration tools.
GitHub issue templates are included for bugs, features, benchmarks, and integrations. Benchmark claims need a reproduction command and committed result artifact before they are added to README.
MIT. See LICENSE.