The CaBrain memory organ as a togo plugin — a
hardware-elastic memory brain for AI agents on togo-postgres (VectorChord + BM25 + pgvector):
hippocampal hot / cortical cold tiers, a Redis L1 cache, salience-gated sleep consolidation,
and reconsolidation-on-recall. Exposed as MCP tools + a Claude Code Memory Tool backend.
togo install togo-framework/brain- retain (§4.1) — embed → recall neighbors → ADD/UPDATE/INVALIDATE/NOOP write decision → compute salience → store as hot episodic → populate the entity graph.
- recall (§4.2) — scoped hybrid retrieval (dense vector + BM25 fused with RRF + a salience nudge) → rerank → 1-hop entity expansion. Hot-tier only; p95 < 300 ms (N1).
- reconsolidation on recall (§4.4) — contradicted facts are superseded, never duplicated.
- Never hard-deletes;
forgetsoft-invalidates and history stays queryable.
brain owns its schema (internal/brain/schema.sql, applied by Migrate) and its queries in Go
— the togo plugin-schema convention. Providers are separate driver plugins registered on the
Service via provider seams (Embedder, Reranker, Engine):
| Provider plugin | Role |
|---|---|
brain-tei |
embeddings + rerank (TEI → Qwen3-Embedding-0.6B, bge-reranker-v2-m3) |
brain-cognee |
cognify engine — entity/graph extraction |
brain-cold-* |
cold-tier demotion (Iceberg/Parquet on S3/MinIO) — Phase 2 |
brain-redis (optional) |
L1 working-memory cache |
Developed in the cabrain project (dev harness);
this repo is the split-out, installable plugin. Design lives in the cabrain project (SPEC.md).
Phase 1: schema + data layer + provider seams in place; retain/recall execution and migrate
land once the provisioned cabrain DB (vchord stack) and the brain-tei provider are live.