LLMSkillHub
mirroredCopied from FerroxLabs/ijfw under MIT. FerroxLabs has not claimed this page, and is not affiliated with LLM SkillHub.

ferroxlabs / ijfw-auto-memorize

Session-end auto-extraction of lessons, errors, fixes, and user feedback into structured memory. Fires at session end. Requires consent on first run.

0
SKILL.md
3.6 KB

Fires at session end. Reads deterministic signals captured during the session
and synthesizes structured memories. Nothing leaves the machine unless the user
explicitly configured an API model via IJFW_AUTOMEM_MODEL.

Consent gate (first run only)

Before any synthesis, check .ijfw/.automem-consent:

  • If missing: ask the user once: "IJFW can automatically extract lessons (errors hit, fixes applied, preferences you stated) at session end into local memory. OK? (y/n). Reply y, n, or ask (ask again next time)." Write answer as {"consented": true|false, "at": "<iso>"} to .ijfw/.automem-consent.
  • If "consented": false: do nothing this session.
  • If "consented": true: proceed.

Inputs (all local files)

  • .ijfw/.session-signals.jsonl -- ERROR/FAIL/Traceback lines captured by the PreToolUse hook (W3.6).
  • .ijfw/.session-feedback.jsonl -- corrections/confirmations/preferences detected by the UserPromptSubmit hook (W3.7).
  • .ijfw/.prompt-check-state -- last turn's intent + vague signals.
  • .ijfw/memory/project-journal.md -- existing entries (dedupe against these).
  • Transcript read via Claude Code's Stop-hook payload (transcript_path).

Synthesis

For each signal cluster:

  1. Redact secrets first. Call redactSecrets() from mcp-server/src/redactor.js on every field that came from transcript or tool output.
  2. Cap sizes. Run applyCaps from mcp-server/src/caps.js. content ≤4KB, why/how ≤1KB, summary ≤120.
  3. Dedupe. Use BM25 search (mcp-server/src/search-bm25.js) against project-journal.md. If score > 6 against an existing entry, skip (duplicate).

  4. Classify into one of:

    • pattern -- error→fix recurrence (same error type seen >=2x).
    • decision -- an explicit user choice ("from now on X").
    • preference -- a style/workflow preference ("I prefer Y").
    • observation -- something worth noting, single instance.

  5. Emit via ijfw_memory_store MCP tool with fields:

    • type: one of the above
    • summary: single sentence, ≤120 chars
    • content: the fact + minimal context
    • why: where this came from (e.g., "user said 'don't use X'", or "hit error Y at step Z")
    • how_to_apply: when this should surface in future sessions
    • tags: include auto-memorize and the classifier kind (correction, confirmation, preference, rule, error)

Model routing

IJFW_AUTOMEM_MODEL env var controls synthesis:

  • unset or off -- skip LLM synthesis; only deterministic signals promoted 1:1.
  • claude-haiku-4-5-* -- Anthropic Haiku (~$0.001/session).
  • ollama:<model> -- local Ollama, fully offline.

Default ship: unset. Deterministic signals still become memories; only the
richer "what did I learn" synthesis is gated on an LLM budget.

Output to user

One-line summary in the terminal:

Stored 3 new memories: pagination-off-by-one fix, user prefers esbuild, stopped repeating rm -rf warnings.

No summary on zero-emit sessions.

Audit trail

Every auto-stored entry carries tags: [..., "auto-memorize"]. The
/ijfw memory audit command lists recent auto-entries for review/removal.

Safety

  • Never store raw transcript content -- only redacted + capped extracts.
  • Never call out to an LLM unless IJFW_AUTOMEM_MODEL is set AND consent is true.
  • Never store secrets -- the redactor runs first, always.
  • Never silently overwrite user-authored memories -- auto-entries go into the knowledge file with their distinguishing tag.

Resume normal mode after.