Recall

Search Clawd's memory for facts related to a query.

hongyime Updated

File contents

/recall — search the knowledge base

Usage:

/recall <query>

Action:

  1. Take everything after /recall and its following space as the search query.
  2. Call the MCP tool kb_recall with:
    • query: the user's query, verbatim
    • k: 5 (the default — bump to 10 if the query is broad, e.g. "what do you know about Caroline")
  3. The tool returns { insights: [{ id, text, source, source_id?, tags?, entities?, category?, importance? }, ...] }.
  4. Format the reply as one fact per bullet, quoting the text field. Cite the source if it's gmail / outlook / photo (the user knows that means it came from ingestion, not from his own messages).
  5. If insights is empty: reply No memories found for "<query>". Try broader keywords or use /memory to teach me first.

Tips:

  • Recall is semantic (pgvector embeddings), not keyword search. Phrase the query the way the user asked, not the way the fact might have been stored.
  • If you suspect the embedding model is offline (recall returns nothing for a query that should obviously hit), call kb_status once to confirm before claiming "I don't remember".

hongyime/pocketclawd commit 461c4397dd

Frequently asked questions

npx get-skill add hongyime/recall