trove-recall

Good answers don't disappear into chat history — they become graph atoms. Match Scribe depth: exact lookup and full-page read first;

recall

is for open questions.

When to use

• A factual question about anything previously ingested or remembered.
• A synthesis ("compare X and Y", "state of Z", "what am I working on").
• A recommendation grounded in the user's own context.

• The answer is in the current working directory's code — read it directly.
• Fresh research with no coverage — research first, then offer

  trove-ingest

  .

Process

Step 1 — pick the retrieval tool (order matters)

grep { pattern }

recall

INV-1042

ECONNRESET

read

billing-pricing-rules

onboarding-checklist

read { slug }

recall { query, tokenBudget }

• The

  recall

  pack is a brief, not always the whole note.
• If the top hit is right but thin →

  read

  that slug.
• Never use

  recall

  for a lone ticket id or error string.

Step 2 — drill down (only if needed)

• read { id | slug }

  for full body + evidence, or a raw source document by source id.

• neighborhood { nodeId, depth, asOf? }

  when structure or belief history matters.

Step 3 — synthesize

• Lead sentence answers the question.
• Cite node titles/slugs for every non-trivial claim.
• Surface gaps explicitly; offer

  trove-ingest

  for missing coverage.
• Surface contradictions with both citations; recommend

  trove-lint

  .

Step 4 — file the answer back (when non-trivial)

• Synthesis across several atoms →

  remember { type: "claim" | "decision" | "pattern", title, summary, evidence, links }

  citing the atoms it drew from.
• Plain fact lookup → nothing to file.

Step 5 — confirm

Anti-patterns

• Don't call

  recall

  for an exact identifier — use

  grep

  .
• Don't answer a known runbook from a thin pack —

  read

  the slug for the full body.
• Don't chain grep→read→read→neighborhood when one well-chosen

  recall

  already answers an open question.
• Don't dump the raw context pack on the user. Synthesize.
• Don't remember trivial lookups back into the graph.