recall

Recall

Multi-layer memory retrieval: structured DB first, markdown fallback, grep sweep.

Workflow

1. Structured search — fast, ranked, typed cells:

   mem search "<query>"
   

   Returns cell ID, type, scene, salience, and content snippet.

2. Markdown memory search — if the platform provides memory_search, use it:

   memory_search(query="<query>")
   

   Then memory_get(path, from, lines) to pull relevant snippets.

3. Grep fallback — catch anything the other layers missed:

   grep -ri "<keyword>" memory/ MEMORY.md
   

4. Synthesize — combine results across layers. Cite sources when helpful:

   • Structured: [cell #ID, scene:<scene>]
   • Markdown: Source: <path>#<line>

Guidelines

• Run all applicable layers (don't stop at the first hit — cross-reference).
• For ambiguous single-word queries, search broadly then ask for clarification if results are thin.
• If nothing is found across all layers, say so clearly.
• When results exist but are sparse, mention what you found and offer to dig deeper.
• Never fabricate memories. Only report what's actually stored.

Storing New Memories

If the recall process surfaces a gap worth filling, offer to store it:

mem quick-store <scene> <type> <salience> "<content>"

Cell types: fact, decision, preference, task, risk, plan, lesson Salience: 0.1 (trivia) → 0.5 (normal) → 0.8 (important) → 1.0 (critical)