memory-reviewer

Files (1)

• SKILL.md

SKILL.md (first lines)

# Memory reviewer

This only reports; it changes nothing. Fixing is `dream`'s job.

Read the curated set with `get_memories` using `exclude_sources: ["session", "archived"]` and `limit` 1000 — raw captures are transcript chunks, not facts, and auditing them is noise. Compare the returned count against the response's `total_matching`: if fewer came back than matched, the audit is **partial** — say so in the header and in the verdict. If the response lacks `total_matching`, the server predates the filter patch; label the audit partial by default.

Group by `metadata.type`. Then scan for:

- **Duplicates** — entries saying the same thing in different words (compare within and across groups).
- **Contradictions** — pairs that can't both be true (e.g. "deploy via CI" vs "deploy by hand").
- **Untyped** — memories with no `metadata.type`.
- **Stale** — facts tied to a tool/version/path that the current repo no longer matches. Flag as *suspected*, not certain; you can't always tell from the text alone.

Output counts first — scanned vs total — then the specific pairs/ids worth a human's attention:

```
review: <scanned> of <total_matching> curated memories — <dups> dup, <contra> contradiction, <untyped> untyped
- dup: [mem0:a1] / [mem0:b2] — both say deploys go through CI
- contradiction: [mem0:c3] vs [mem0:d4]
```

End by pointing at `dream` if there's anything to merge or prune.

From pjope/mem0-selfhosted-hooks, MIT.

More from pjope/mem0-selfhosted-hooks