| name | learnings |
| description | Review the current session for learnings worth persisting. Prefers committed files (CLAUDE.md, skills) over memory — memory is only for personal/user-specific things. |
| allowed-tools | Read Write Edit Glob |
Review this session and persist any learnings that will be useful in future conversations.
Where learnings go — the hierarchy
Learnings should be shared with the whole team by default. Only use memory for things that are genuinely personal. Apply this decision tree:
- Project convention or workflow rule → update
CLAUDE.md or the relevant skill in .claude/skills/. These are committed to the repo and shared with all contributors.
- Skill-specific guidance (how to write tests, how to format commits) → update the relevant skill file directly.
- Personal preference (user's communication style, role, expertise) → save as memory. These are per-user and not shared via the repo.
- Project context (deadlines, who's doing what, external references) → save as memory only if it's not derivable from code, git, or docs.
Rule of thumb: if a future contributor using this repo would benefit from the learning, it goes in a committed file, not memory.
What to look for
Scan the full conversation history for:
- User corrections — "don't do X", "use Y instead", "that's wrong", "I prefer Z".
- Confirmed approaches — things that worked well and the user approved, especially non-obvious choices.
- Project context — decisions, deadlines, goals, why something is being done.
- User profile updates — new info about the user's role, expertise, or preferences.
- External references — URLs, tools, dashboards, ticket systems mentioned.
What NOT to save
- Anything already in CLAUDE.md, skills, or derivable from code/git
- Ephemeral task details (what files were edited, what commands ran)
- Debugging solutions (the fix is in the code)
- Things that are only relevant to this session
Process
- Read
CLAUDE.md and skim the existing skills to know what's already documented.
- Read
MEMORY.md at the memory path to see what's already recorded.
- Identify candidate learnings from the session — list them briefly.
- For each candidate, classify using the hierarchy above:
- Committed file update → apply the change to CLAUDE.md or the skill file directly.
- Memory → write/update the memory file and MEMORY.md index.
- Skip → already covered or too ephemeral.
- Report what was done.
Output format
Summarize what you did:
- Changes applied to committed files (CLAUDE.md, skills) with one-line descriptions
- Memories added/updated/removed (if any)
- Nothing saved? Say so — not every session produces learnings
Keep it brief — the user invokes this at session end and wants a quick summary, not an essay.
$ARGUMENTS