| name | commit-learnings |
| description | End-of-session sweep — review what you learned about the user during the conversation and commit durable facts to the configured personal-memory backend. Use when wrapping up a session, when the user says "save what you learned" or "commit this to memory", or proactively before a handover. Complements remember-user-fact (which saves in-the-moment); this skill catches facts that slipped through. Concrete backend (Pinecone, Mem0, …) comes from the workspace's .claude/memory-config.md. |
Commit end-of-session learnings
Over a long session, facts accumulate that weren't important enough to trigger remember-user-fact in the moment but, in aggregate, amount to a meaningful update to what memory knows about the user. This skill is the sweep — a structured review before the session ends.
When to invoke
- User says "save what you learned", "commit this to memory", "update memory", or similar.
- User is wrapping up a session (handover,
/clear, end-of-day).
- Proactively, when a long session has clearly surfaced durable facts you haven't saved yet.
- Before a handover document is written — the next session should inherit the learnings.
Do not invoke:
- In short sessions where nothing new was learned about the user.
- When the only "learnings" are transient task state (what got done this session belongs in a handover doc, not memory).
How to run it
1. Load the config
Read .claude/memory-config.md in the workspace. You will need the backend's search tool (for the duplicate check) and add/update tool (for the writes), plus scope parameters for both personal and work contexts. If the file is missing, stop and ask the user to install one.
2. Scan the session
Walk back through the conversation and list every candidate fact. For each, ask:
- Is it about the user (preferences, role, context, corrections) rather than the code or task?
- Is it durable — still true next week, next month?
- Is it not already in memory? Quickly search to check.
- Is it not sensitive (no credentials, tokens, private third-party info)?
A fact that passes all four is a commit candidate.
3. Group by context
Separate candidates into personal and work piles using the deduction rule in CONTEXT.md. A single session can produce commits to both stores — that's fine, just keep them separate.
4. Show the user before saving
This skill is different from remember-user-fact because it's batch and retrospective — the user may not remember telling you half of this. Before saving, show the proposed commits:
I'd like to save the following to memory:
PERSONAL (2):
• <fact 1>
• <fact 2>
WORK (1):
• <fact 3>
Proceed? (you can edit, skip items, or reclassify)
Respect edits and skips. Never save a fact the user rejected.
5. Save
For each approved fact, call the configured add tool with the right scope parameters (see remember-user-fact SKILL.md for the per-record construction). Before each save, do a quick duplicate check via the configured search tool; if a close match exists, prefer an update over creating a duplicate.
6. Report
One short summary line: "Saved 2 to personal, 1 to work." Do not re-list the facts.
Quality bar
- Atomic commits. One fact per memory. Do not concatenate the whole session's learnings into one blob.
- Be stingy. A session that saves 10 memories is usually a session with 8 false positives. Three well-chosen commits beat ten noisy ones.
- If in doubt, skip. You can always save a fact next session when it comes up again. A missing memory is cheaper than a wrong one.
