joshuaswarren/remnic/remnic-recall

When to use

Use when the user or the current task needs prior context from Remnic. This is the default first step for any non-trivial turn that could benefit from memory.

Triggers:

• "What do you remember about …"
• "Have we talked about …"
• "Recall anything on …"
• A new task begins and the agent wants background.

Inputs

• query (required) — natural-language question or topic string.
• Optional budget hint from the caller (e.g., "brief", "deep").

Procedure

1. Build a concise natural-language query from the user's request. Prefer the user's own wording over paraphrase.
2. Call remnic_recall with that query. Ask for 3–8 results unless the caller hinted otherwise.
3. Skim the returned memories. Discard anything clearly off-topic.
4. Present 1–5 relevant bullet points to the user, each attributed to its source memory when useful.
5. If nothing relevant came back, say so plainly and suggest remnic-remember if there is something worth storing now.

Efficiency plan

• One broad recall beats several narrow ones.
• Reuse results within the same turn — do not re-query for the same topic.
• Skip recall entirely for trivially local requests (formatting, arithmetic, mechanical refactors).

Pitfalls and fixes

• Pitfall: Quoting irrelevant recalls just because they came back. Fix: Filter by topical relevance before surfacing.
• Pitfall: Over-narrowing the query and missing useful context. Fix: Start broad; refine only if the first pass was noisy.
• Pitfall: Presenting raw memory blobs. Fix: Summarize in the user's own terms.

Verification checklist

• [ ] remnic_recall was called with a natural-language query.
• [ ] Results were filtered for relevance before surfacing.
• [ ] User-facing summary is concise (≤ 5 bullets unless requested).
• [ ] Legacy engram_recall alias was not preferred over remnic_recall.

Tool names: canonical name is remnic_recall. The legacy engram_recall alias remains accepted during v1.x.