reflect-and-savelisted

# Reflect & save (self-improving loop) After a meaningful task, reflect and persist what's durable. The model isn't retrained — it keeps better notes. ## Steps 1. Ask: did anything happen worth remembering next session? - user correction/preference (how they want you to work) -> type: feedback - non-obvious project/infra fact NOT in code or git -> type: project - a decision + its reason -> type: project - pointer to an external resource -> type: reference - who the user is -> type: user 2. If NOTHING durable -> say so and stop. Don't save noise or facts the repo/git already record. 3. If yes -> write ONE memory file per fact (frontmatter: name, description, metadata.type), update the MEMORY.md index. Check for an existing file first -> update, don't duplicate. Link related memories with [[name]]. 4. If a workflow recurred, capture it as a new Skill instead of a memory. 5. Keep entries compact. ## When to skip Trivial edits, pure conversation, or anything already captured. ## v2.3 additions - First check `~/.claude/reflect-staging.md` (auto-captured failures) and curate anything durable from it. - PROVENANCE GUARD: never paste GLM/M3/Codex text verbatim into memory — rewrite the lesson in your own words, human-confirmed. - Prefer rewriting an existing memory/lesson over appending a near-duplicate (avoid bloat).