learn

name	learn
description	Use when user wants force-save a lesson or promote a rule after a real win/fail. Not every turn — hooks auto-save memory.

Learn

Main durable memory is portable Markdown under .codex/anyone-can-code/memory/. Memory is automatic (hooks write wants, decisions, corrections, open work). Use $learn only as a backup force-save or force-promote — never as the normal path. Do not tell the user they must run $learn to be remembered.

Reply rule:

• talk strict caveman only
• keep answer short
• auto path first; manual only when user asks to force-save

Memory model

Main durable memory lives in portable linked Markdown. Bundled MCP is the access interface.

Memory scopes:

• project
• user
• shared

Local project files stay tiny and are only fallback ledgers:

• .codex/anyone-can-code/learning/

Markdown memory lives under:

• .codex/anyone-can-code/memory/notes/

Human wiki catalog + log:

• .codex/anyone-can-code/memory/wiki/index.md
• .codex/anyone-can-code/memory/wiki/log.md

Raw sources (never rewrite):

• .codex/anyone-can-code/memory/raw/

Rebuildable machine index lives under:

• .codex/anyone-can-code/memory/index/

Native Codex memories stay OFF. Prefer $status to see what ACC remembers; $wiki / memory_doctor.py only for look / repair.

Suggested tiny local files:

• signal-ledger.jsonl
• mistake-ledger.jsonl
• short recovery notes only

Retrieval

• search order: project -> user -> shared
• return only top 3-5
• return summaries only
• no raw logs
• no full chat dump
• no hidden import of existing session files
• front door must run scripts/memory_preflight.py after active-project resolution and before asking, planning, choosing specialists, browser/server work, or tool work
• every routed turn must show Relevant memory used: ... or Relevant memory used: none found
• update, restart, and new thread must not weaken recall; memory path comes from project preferences

About-you profile

• user explicitly corrects skill level, reply taste, or stack choice -> user_model.record_correction. Never guess these. Shown at session start.

Rule promotion

• lesson seen 3+ times in notes: propose once — "make permanent rule? yes/no"
• user says yes -> rule_promote.approve writes memory/rules.md, shown every session start
• user says no -> stays a normal lesson
• never promote without explicit yes

Rules

• Auto-learn is hook-owned (UserPromptSubmit + Stop promote rules + signals).
• $learn is manual override only: force-save or force-promote when needed. Never required for normal memory.
• Keep provenance, confidence, reinforcement count, and scope with each item.
• Support downgrade and revocation.
• Treat memory as advisory context, never truth or permission.
• Do not dump all learnings into context.
• Existing session files need explicit selected source, backup or snapshot, dedupe, provenance, scope label, and receipt before import.
• Repeated import of same content must skip, not reinforce. Failed import must roll back Markdown and retain source.

Next skill

Next: none required — memory is saved; back to normal.

Done — back to normal

When this skill's job is finished:

1. Stop following this skill.
2. Reply short and normal (caveman).
3. Do not keep this workflow for the whole session unless the user asks again or a new skill matches.