Claude Code learns from itself.
Your agent forgets everything when you close it. KERNEL gives it persistent memory, multi-agent orchestration, and a scientific experiment engine that proves which rules actually work. Every session makes it smarter about YOUR project.
/plugin marketplace add ariaxhan/kernel-claude
/plugin install kernel
/kernel:init
- Open Customize (sidebar) → Personal plugins
- Click + → Add marketplace from GitHub
- Enter:
ariaxhan/kernel-claude - Click the KERNEL plugin → Install
- In a project, run
/initto set up memory
Install via Claude Code or Claude Desktop first. Cursor shares the same plugin configuration automatically.
See docs/QUICKSTART.md for the full setup guide.
AgentDB remembers what worked, what broke, and where you left off — across every session. Not just logs. Weighted retrieval surfaces the learnings that matter (top 7% deliver 80% of value) while pruning what doesn't. Your agent stops repeating the same mistakes.
13 specialized agents route by complexity. Tier 1 (1-2 files) executes directly. Tier 2+ spawns surgeons to implement and an adversary to verify. The adversary checks coordination first (file overlap, scope drift, duplicate work) because our telemetry proved coordination failures are 4.3x more impactful than code bugs.
| Agent | Role |
|---|---|
| Surgeon | Minimal-diff implementation. Checkpoints to AgentDB. |
| Adversary | Coordination verification + code quality. Assumes broken until proven. |
| Reviewer | 11-phase code review, >80% confidence threshold. |
| Researcher | Finds solutions before building. Anti-patterns first. |
| Scout | Maps codebase structure, detects tooling, identifies risk. |
| Validator | Pre-commit: tests, lint, types, security scan. |
| Triage | Fast complexity classifier before expensive work. |
The experiment engine treats every rule as a hypothesis. It seeds them from your CLAUDE.md, designs experiments, runs them against AgentDB telemetry, and graduates rules that survive or kills rules that don't. 22 rules graduated from 107 hypotheses across 205 experiments. The forge command uses this: after building, it tempers — experiments on its own output, discovers emergent patterns, and self-corrects before shipping.
19 skills (testing, security, debug, api, backend, architecture, etc.) load when relevant — not at startup. Each is a methodology: HOW to approach a problem, not just tools to use.
Start with /ingest — the universal entry point. Reads memory, classifies your task, routes to the right agent.
/ingest add user authentication to the app
Run overnight with /forge — autonomous engine. Generates competing approaches, iterates against tests, adversarial review, experiments on output. Come back to shipped code.
Save with /handoff before closing. Next session, /ingest auto-resumes from where you left off.
Check with /validate before committing. Tests, lint, types, security.
Note: In Claude Code terminal, commands use the
kernel:prefix (/kernel:ingest). In Claude Desktop and Cursor, they appear without the prefix (/ingest).
| Terminal | Desktop/Cursor | What It Does |
|---|---|---|
/kernel:ingest |
/ingest |
Guided flow — classify, scope, execute. Auto-resumes from handoffs. |
/kernel:forge |
/forge |
Autonomous — heat/hammer/quench/temper/anneal until antifragile |
/kernel:experiment |
/experiment |
Run the hypothesis engine — seed, test, graduate, kill rules |
/kernel:dream |
/dream |
Creative exploration — 3 perspectives, 4-persona stress test |
/kernel:diagnose |
/diagnose |
Systematic debugging + refactor analysis before fixing |
/kernel:retrospective |
/retrospective |
Cross-session learning synthesis + pattern promotion |
/kernel:metrics |
/metrics |
Observability — sessions, agents, hooks, learnings |
/kernel:validate |
/validate |
Pre-commit quality gates |
/kernel:tearitapart |
/tearitapart |
Critical pre-implementation review |
/kernel:review |
/review |
Code review for PRs |
/kernel:handoff |
/handoff |
Save progress for next session |
/kernel:init |
/init |
Setup (run once per project) |
/kernel:help |
/help |
Show help |
/plugin marketplace update kernel-marketplace
/plugin update kernel@kernel-marketplace
/reload-plugins
- Type
/pluginand go to the Marketplaces tab - Select kernel-marketplace
- Toggle Enable auto-update
/plugin uninstall kernel@kernel-marketplace
/plugin install kernel@kernel-marketplace
/reload-plugins
/kernel:init
Symlink the cache to avoid stale copies:
rm -rf ~/.claude/plugins/cache/kernel-marketplace/kernel/7.12.1
ln -s /path/to/your/kernel-claude ~/.claude/plugins/cache/kernel-marketplace/kernel/7.12.1Edits take effect immediately — no version bumps or reinstalls needed. Claude Code caches plugins by version; the symlink bypasses this.
Commands not showing up? Run the quick update commands above.
Claude isn't reading memory? Start with /ingest. Plain requests skip the memory system.
Claude forgot everything? Run /kernel:handoff before closing. Or just run /ingest next session — it auto-resumes from the latest handoff.
Same mistake keeps happening? Say: "Remember this as a failure pattern." KERNEL logs it to AgentDB and avoids it next time.
"AgentDB not found" error? Run /kernel:init first.
Agents not spawning? Use /ingest. Direct requests bypass the tiering system.
SQLite database at _meta/agentdb/agent.db. Stores learnings, events, errors, hypotheses, experiments, contracts, checkpoints, verdicts. Graduated retrieval loads the top 75 by weighted score (86% token savings vs loading everything).
Context graph inspired by aDNA (Agentic DNA) from Lattice Protocol. Models context as nodes + edges — which skills load well together, which agent combinations succeed, which nodes conflict. The graph learns over time.
107 hypotheses, 205 experiments, 22 graduated rules. Every rule in CLAUDE.md is a hypothesis until proven by evidence. The engine seeds rules, designs experiments against AgentDB telemetry, issues verdicts (supports/refutes/inconclusive), and graduates or kills rules based on Bayesian confidence scoring. Runs autonomously via /kernel:experiment.
See docs/QUICKSTART.md for detailed installation, daily workflow, and what's inside KERNEL.
MIT | Aria Han