Memory Wrap

Turn the durable part of this conversation into memory. The bar: would knowing this in six months change how the user works or decides? If yes, propose it. If it is marginal, skip it. This is not a log of the conversation.

All tools below are on the imi MCP server (possibly prefixed, e.g. mcp__imi__capture_thought). If no imi tools are available, say so and stop.

1. Check what imi already knows

Run 2–3 memory_recall queries on the main subjects, people and projects of the session, with record_kinds ["capture", "agent_memory"]. Do not read everything — just enough to avoid proposing what is already stored. Keep each request_id.

2. Pick candidates (be skeptical)

• Decision — a concrete choice about the work, its direction or process.
• Fact — specific, non-obvious, new information about people, organizations, systems, commitments.
• Lesson — an insight, a named concept, a way of framing something that clicked.

Skip anything already stored, process chatter without a conclusion, things that only matter inside this conversation, and anything vague. Never propose secrets, credentials, access tokens, or large pasted code or transcripts.

Aim for 3–7. If you have 10+, tighten the filter. If there is nothing, say so — not every session produces durable knowledge.

3. Ask before saving

Found 4 things worth saving — "all", "none", or numbers (e.g. "1, 3"):

1. [DECISION] <one self-contained sentence> · tags: a, b, c
2. [FACT] ...
3. [LESSON] ...

Wait for the answer. Save nothing without it.

4. Save what was approved

Choose the tool by shape:

• A few standalone items — one capture_thought per item: content = the item, prefixed with its type ([DECISION], [FACT], [LESSON]) and expanded enough to stand alone months from now; tags = the proposed tags plus the type; source = "manual". deduped: true means imi already had it — nothing more to do.
• An end-of-task batch (the session finished a defined piece of work) — one memory_writeback with memory_payload holding decisions, lessons, next_steps, unresolved_questions as lists of strings, and an idempotency_key that is stable for this session (e.g. the date plus a short slug), so a retry does not duplicate.

What you save is evidence, pending review (ADR-002): searchable immediately, but it only becomes instruction-grade when the user confirms it in imi's review queue. The user's "yes" here decides what is written, not what is trusted — say so in one line when reporting. Never try to mark memories confirmed yourself.

5. Report and close the loop

One line: how many saved, how many already present. Then call record_memory_usage for each recall request_id, listing the record ids that changed what you proposed or skipped.

Tone

Housekeeping, not ceremony. The list should take ten seconds to review.