memory-digest git:20260510.7af9308A

Immutable. This exact content never changes and is served at /api/v1/blob/5b789db7ef645cde.

---
name: memory-digest
description: Digest recent tape anchors and session context into curated pi-memory-md memories when tape mode is enabled. Use only with tape mode enabled when the user asks to summarize recent work, extract learnings, preserve decisions, convert tape/session history into memory, or review recent intent for durable memory updates.
---

# memory-digest

Turn recent tape anchors and nearby session context into concise, durable memory. Prefer curation over transcript summarization.

## Core rule

Inspect recent intent first, read only relevant context, propose memory changes, then write only after user confirmation using the `memory-write` skill.

## Workflow

1. Inspect recent anchors with `tape_list`, usually `limit: 20` and `contextLines: 1`.
2. Infer the user's recent intent from handoff anchor names, summaries, purposes, keywords, timestamps, and nearby context.
3. Select the smallest useful tape range by combining anchor count and recency:
   - Use `tape_search({ kinds: ["anchor"], anchorType: "handoff", anchorScope: "project", betweenDates: { start, end }, limit: 20 })` for handoff anchors from the last 2 days.
   - Use `tape_search({ kinds: ["anchor"], anchorType: "handoff", anchorScope: "project", limit: 10 })` when recent date-filtered handoff anchors are not enough to capture the active thread.
   - Use `tape_read({ lastAnchor: true })` for the latest active thread.
   - Use `tape_read({ afterAnchor })` when one anchor starts the relevant work.
   - Use `tape_read({ betweenAnchors })` when a clear start/end pair exists.
4. Extract durable memory candidates only when they are useful across future sessions.
5. Compare candidates with existing memory using `memory_list` or `memory_search` before proposing creates or updates.
6. Present proposed changes and ask for confirmation.
7. After confirmation, use the `memory-write` skill to create or update memory files.
8. Ask before `memory_sync` unless the user explicitly requested sync.

## Save candidates

Save concise knowledge in these categories:

- user preferences
- project decisions
- architecture constraints
- reusable implementation patterns
- recurring bugs and fixes
- workflow habits
- unresolved questions or future directions

Do not save:

- raw command output
- temporary debugging noise
- one-off implementation details recoverable from git
- unconfirmed guesses
- large chat summaries
- secrets, credentials, tokens, or `.env` content

## Proposal format

Before writing, respond with:

```md
## Proposed memory updates

### Create
- `path/to/file.md`
  - reason:
  - summary:

### Update
- `path/to/file.md`
  - reason:
  - summary:

### Skip
- item:
  - reason:
```

End by explicitly asking the user for confirmation with a direct question, for example:

`Proceed with these memory updates?`

Do not write memory files until the user confirms.

## Proposal guidance

During analysis, propose likely memory paths and metadata, but treat them as suggestions for user confirmation and final `memory-write` handling.

- Propose paths based on the candidate's scope, such as project direction, reusable reference, research note, or archive material.
- Keep each proposed memory focused on one durable topic.
- Propose a concise description and 2-5 lowercase reusable tags for each candidate.
- For possible updates, identify the existing memory file and explain what should be merged.
- Defer final placement, frontmatter shape, timestamp handling, and file creation/update mechanics to the `memory-write` skill.

## Related skills

- `memory-write`: Create or update memory files with valid frontmatter.