name	cited-recall
description	Answer a question from memory with per-claim citations, GBrain-style — written answer, sources for every claim, explicit about what the memory does not contain. Use when the user needs a sourced or auditable answer about past sessions/decisions.

/cited-recall

Answer a question from the memory store with a synthesized, cited answer instead of raw chunks. This is the GBrain pattern (reranker + cited prose + "what the brain doesn't know") rebuilt on memsearch primitives with Claude as the reranker — zero extra infrastructure, zero API cost beyond the session itself.

Inputs

A question about past sessions, decisions, people, or project history.

Procedure

Freshness. Run memsearch index .memsearch/memory/ first (the index lags mid-session; see CLAUDE.md "Index freshness").
Retrieve (L1). Run memsearch search "<question>" — and at least one rephrased variant if the first pass looks thin. Collect the top ~5 candidate chunks.
Rerank (Claude-as-reranker). Judge each candidate for actual relevance to the question. Discard near-misses; order the survivors by how directly they answer. Do not trust the similarity score alone — a 0.9 on the wrong entity is still wrong.
Expand (L2). For each survivor, run memsearch expand <chunk-id> to get the full markdown section and its anchors (source file, heading, session id, transcript path).
Drill down (L3, only if needed). If the summaries conflict or lack a critical detail, use memsearch transcript against the anchored transcript to check the original exchange.
Synthesize. Write a direct answer in prose. Every factual claim carries a citation: (memory: <file> § <heading>) — add , transcript <session-id-prefix> when L3 was used. If this repo contains other note collections (people notes, project notes), consult and cite them the same way.
Admit gaps. End with a "Not in memory:" line listing the parts of the question the store could not answer — name what's missing rather than papering over it. If nothing relevant was found at all, say exactly that and stop; never substitute plausible guesses.

Output

A short written answer: cited claims, then the "Not in memory:" line (omit only when the answer is complete). No raw chunk dumps.

Rules

Never present an uncited claim as remembered fact.
A confident answer with no source is worse than "not found".
Prefer fewer, well-supported claims over coverage.

/cited-recall

Inputs

A question about past sessions, decisions, people, or project history.

Procedure

Freshness. Run memsearch index .memsearch/memory/ first (the index lags mid-session; see CLAUDE.md "Index freshness").

Retrieve (L1). Run memsearch search "<question>" — and at least one rephrased variant if the first pass looks thin. Collect the top ~5 candidate chunks.

Rerank (Claude-as-reranker). Judge each candidate for actual relevance to the question. Discard near-misses; order the survivors by how directly they answer. Do not trust the similarity score alone — a 0.9 on the wrong entity is still wrong.

Expand (L2). For each survivor, run memsearch expand <chunk-id> to get the full markdown section and its anchors (source file, heading, session id, transcript path).

Drill down (L3, only if needed). If the summaries conflict or lack a critical detail, use memsearch transcript against the anchored transcript to check the original exchange.

Synthesize. Write a direct answer in prose. Every factual claim carries a citation: (memory: <file> § <heading>) — add , transcript <session-id-prefix> when L3 was used. If this repo contains other note collections (people notes, project notes), consult and cite them the same way.

Admit gaps. End with a "Not in memory:" line listing the parts of the question the store could not answer — name what's missing rather than papering over it. If nothing relevant was found at all, say exactly that and stop; never substitute plausible guesses.

Output

A short written answer: cited claims, then the "Not in memory:" line (omit only when the answer is complete). No raw chunk dumps.

Rules

Never present an uncited claim as remembered fact.

A confident answer with no source is worse than "not found".

Prefer fewer, well-supported claims over coverage.