satori

A self-learning loop for Claude Code — the model learns skills from its own sessions. Fully automatic, fully visible, fully reversible.

English · Русский

⚠️ Beta disclaimer

This is beta software I build for myself and share with the community as-is. It parses your session transcripts, stores lesson data locally and — by default, automatically — activates skills that will instruct your future Claude sessions. Every activation is announced in chat (⛩) and reversible in one call (retire_skill), but do read server.py and watch what your agent learns. It works on my setup; I can't guarantee yours and take no responsibility for the skills your agent teaches itself. Want to pre-approve everything instead? SATORI_AUTO_APPROVE=0.

satori is an MCP server + hooks that give Claude Code a closed self-learning loop: user corrections and tool failures become lesson candidates, lessons become skill drafts, drafts become active skills — automatically, visibly, reversibly. Windows-native, works in Claude Desktop, zero bash wrappers.

Two principles you won't find together elsewhere: the server does only deterministic mechanics (parsing, counters, storage, validation) — the thinking is done by the calling model right in the session, no background LLM calls, no extra bills; and fully automatic, fully visible, fully reversible — validated drafts activate on their own, every loop event is announced in chat with a ⛩ marker, and retire_skill undoes any activation in one call (prefer a manual approval gate instead? SATORI_AUTO_APPROVE=0).

Features

• 4-stage loop — capture → decide → distill → curate; reflect() is called several times per session

• User correction = signal #1 — surfaces as a candidate after a single occurrence (Devin's mechanic); tool failures wait for recurrence

• Patch-not-append — a recurring signal bumps seen_count instead of piling up rows

• SKIP gate, forever — "not skill-worthy" verdicts are remembered; rejected noise never comes back

• Fully automatic — a validated draft activates immediately (backup of anything it overwrites, retire_skill = one-call undo); SATORI_AUTO_APPROVE=0 switches to a manual staging gate

• The trigger is sacred — description: Use when ... is mandatory: recall works when a note says when to recall it, not what it does

• pinned_project — a lesson is global or pinned to one project; approve routes it to the right skills folder

• Draft validation — frontmatter, size, secrets (also redacted at rest), prompt-injection markers (EN+RU)

• FTS5 search over past sessions — "when did I fix exactly this error" finds the actual transcript

• Curator — usage telemetry; unused drafts go stale in 30 days, archived in 90

• Smart nudge hooks — silent by default; they speak only on a correction (instantly) or accumulated work; a declined nudge stays declined

• Visible audit trail — every loop event surfaces in chat as a ⛩ marker line: what fired, why, what was staged or skipped

• dream/wake integration — the dream-skill consolidation pass harvests satori's staging and promotes/retires drafts through its validator gate

Related MCP server: Claude Error Collector

Quick Start

Easiest — let Claude install it. Paste this message into Claude Code:

Install the satori self-learning loop from https://github.com/timoncool/satori:
1) clone the repo to a permanent location (not a temp dir — the MCP runs from it);
2) create a venv inside it and install the single dependency: fastmcp;
3) register the MCP in my .mcp.json (project or global) as "satori" with
   command = absolute path to the venv python and args = [absolute path to server.py];
4) recommended: add the three nudge hooks (UserPromptSubmit / Stop / SessionEnd,
   all calling hooks/nudge.py with the venv python — exact JSON is in the README
   "Quick Start" step 3) into ~/.claude/settings.json, PRESERVING my existing hooks;
5) smoke-test: import server in the venv and show me what got configured;
then remind me to restart Claude Code / Desktop so the MCP and hooks load.

That's it — Claude clones, wires configs, verifies and reports. Manual way:

1. Clone & install the one dependency

   git clone https://github.com/timoncool/satori.git
   cd satori && python -m venv .venv && .venv\Scripts\pip install fastmcp

2. Register the MCP — in your project's .mcp.json (or global config):

   "satori": {
     "command": "<path>\\satori\\.venv\\Scripts\\python.exe",
     "args": ["<path>\\satori\\server.py"]
   }

3. (Optional) nudge hooks — three hooks onto one script in ~/.claude/settings.json:

   "UserPromptSubmit": [{"matcher": "", "hooks": [{"type": "command", "command": "<venv-python> <path>/hooks/nudge.py prompt-submit", "timeout": 10}]}],
   "Stop":             [{"matcher": "", "hooks": [{"type": "command", "command": "<venv-python> <path>/hooks/nudge.py stop", "timeout": 10}]}],
   "SessionEnd":       [{"matcher": "", "hooks": [{"type": "command", "command": "<venv-python> <path>/hooks/nudge.py session-end", "timeout": 60}]}]

   Restart Claude Code / Desktop.

How it works

session transcript
      │  (hook/tool call — parsing costs 0 tokens)
      ▼
① capture   reflect() reads what's new since the stored offset: corrections,
            failures, fix-after-fail, complex segments (≥12 calls + ≥2 edits)
      ▼
② decide    the model judges candidates (corrections after 1×, the rest after 2×):
            noise → skip_lesson (forever), worthy → a draft
      ▼
③ distill   submit_draft → validated, provenance-stamped, staged AND auto-activated
            into ~/.claude/skills/ (or the pinned project). ⛩ announced in chat;
            retire_skill = one-call undo. Manual gate: SATORI_AUTO_APPROVE=0
      ▼
④ curate    usage telemetry, stale at 30d, archive at 90d

Hooks (all optional, all silent by default): UserPromptSubmit — on a correction injects one line "fix it, then reflect" (series-deduplicated); on ≥25 accumulated tool calls — same, and repeats only after the next full threshold; Stop — same threshold at turn end; SessionEnd — silent capture directly in Python, no model involved at all. When a nudge does fire, the model opens its reply with a visible ⛩ satori: ... marker and reports what was recorded — you always see the loop working.

Anti-pollution by construction: the loop never writes into Claude's memory (only its own SQLite + staging); context injections are one line and only when warranted; an ignored nudge doesn't nag.

Tools

Configuration (env)

Works best with dream-skill

dream-skill is this project's sibling — memory consolidation for Claude Code (dream = read-only pass, wake = gated apply, full rollback). satori handles procedural memory (skills), dream/wake handles factual memory (notes, rules, index) — and they meet in the middle:

• satori runs fully automatic inside sessions; dream is the periodic auditor: its Skill harvest phase reads satori's staging + usage telemetry and proposes retire_skill for stale/duplicate self-taught skills (and promote_skill if you run satori in manual mode)

• dream's validator re-checks every self-taught skill against a hard checklist: Use when trigger, no injection markers, no duplicates

• wake applies the audit, logs everything, and its rollback restores skills too

Each works standalone; together the cycle is complete: сон → пробуждение → прозрение (dream → wake → satori).

Standing on shoulders

An honest list of sources: Hermes Agent (Nous) — the 4-stage cycle, FTS5, periodic self-reflection; claude-self-improving-skills — complexity thresholds, patch-over-create, curator, telemetry, "declined stays declined"; claude-evolve — objective signals, patch-not-append; claude-harness-hermes — permanent SKIP, redaction at rest; Devin (Cognition) — corrections at threshold 1, the trigger as a sacred field, pinned scoping, injection scanning; dream-skill (ours) — the staging gate.

Other Projects by @timoncool

Authors

• Nerual Dreming — Telegram | neuro-cartel.com | ArtGeneration.me

Support the Author

I build open-source software and do AI research. Most of what I create is free and available to everyone. Your donations help me keep creating without worrying about where the next meal comes from =)

All donation methods | dalink.to/nerual_dreming | boosty.to/neuro_art

• BTC: 1E7dHL22RpyhJGVpcvKdbyZgksSYkYeEBC

• ETH (ERC20): 0xb5db65adf478983186d4897ba92fe2c25c594a0c

• USDT (TRC20): TQST9Lp2TjK6FiVkn4fwfGUee7NmkxEE7C

Star History

License

MIT