Sinapsis Learning

Sinapsis Learning v4.1

Continuous learning engine for Claude Code. Observes sessions via deterministic hooks, detects error patterns at session end, and injects learned instincts into future sessions. This skill is ALWAYS ACTIVE — it loads at session start.

How It Actually Works

Sinapsis Learning is built on 4 deterministic bash hooks (no LLM in the pipeline):

PreToolUse  → observe.sh pre   (async) — logs tool + input to observations.jsonl
PostToolUse → observe.sh post  (async) — logs tool output + is_error flag
PreToolUse  → _project-context.sh (sync, once/session) — injects last context.md
PreToolUse  → _instinct-activator.sh (sync) — injects matched instincts
Stop        → _session-learner.sh — writes context.md + detects error patterns
On-demand   → _dream.sh — 5-module index hygiene (duplicates, contradictions, staleness, triggers, health)

What fires automatically: passive rules, instinct injection, observation logging, context.md writing. What requires your input: reviewing proposals (/analyze-session), accepting new instincts, running /evolve.

The Learning Pipeline

You work on a project
        |
        v
observe.sh logs every tool use → observations.jsonl
        |
        v
Session ends (Stop hook)
        |
    _session-learner.sh runs:
        ├── Writes context.md (project name, date, files touched, gotcha count)
        └── Detects error→fix patterns → _instinct-proposals.json (draft)
        |
        v
Next session starts
        |
    _project-context.sh injects context.md (once)
        |
        v
You run /analyze-session
        |
    Review proposals → accept/reject
        |
        v
Accepted instincts → _instincts-index.json (confirmed)
        |
        v
Future sessions:
    _instinct-activator.sh matches instincts → injects as systemMessage

Instinct Levels

Domain dedup: one instinct per domain fires per tool use. permanent beats confirmed. Max 3 domains total.

Instinct Format

{
  "id": "unique-id",
  "domain": "security|git|code-quality|deployment|testing|workflow|...",
  "level": "confirmed",
  "trigger_pattern": "regex matched against tool_name + tool_input",
  "inject": "The message injected as systemMessage when trigger matches.",
  "origin": "manual | learned",
  "added": "2026-01-15"
}

What Gets Observed

Every tool use is logged to ~/.claude/homunculus/projects/{hash}/observations.jsonl:

• Tool name and key input fields (file_path, command, pattern)
• Output excerpt (scrubbed of secrets)
• is_error: true flag when output contains error keywords (error, failed, exception, traceback)
• Timestamp and session ID

Observations stay local — no external transmission, no cloud sync.

Active Capture (On Request)

You can also create instincts manually:

• "Learn this pattern" — Claude creates an instinct immediately (level: confirmed)
• "Never do X again" — anti-pattern instinct
• "Always use Y for Z" — preference instinct

Use /instinct-status to see all instincts and their levels.

Integration with /evolve

When instincts cluster around a theme, /evolve lets you promote them:

• [S]kill: Create a reusable skill from a cluster of instincts
• [C]ommand: Create a slash command for a repeated workflow
• [A]gent: Create an autonomous agent for complex patterns
• [R]ule: Convert to a passive rule (fires on trigger, no instinct index needed)
• [E]nrich: Add knowledge to an existing skill
• [P]romote: Move from project scope to global scope
• [X]: Skip — not ready yet

Privacy

• All data is local (~/.claude/homunculus/)
• No personal data captured in instincts — only patterns
• Delete any instinct: remove from _instincts-index.json
• Delete observations: remove the project's observations.jsonl
• /instinct-status shows everything that has been learned

Commands