mahmoudimus/memories-recall-verify

Self-Correcting Recall

Use this when you are about to answer a question whose answer depends on stored memory (preferences, decisions, facts, "what did we decide about X"), and the memories you have are ambiguous, conflicting, or incomplete. The goal is to ground the answer in the right memory — not the first plausible one — and to admit when memory doesn't contain the answer.

Step 1 — Recall

simba memory recall "<the user's question, as a statement>"

Each line is: <id> [<TYPE>] (<similarity>) <content>.

Step 2 — Detect a problem

Inspect the recalled set and decide whether you can answer directly. You cannot yet if any of these hold:

• Ambiguous reference — the question names a generic thing ("the API key", "the staging DB", "the meeting") and the memories describe multiple distinct instances of it. You must not blend facts across instances.
• Conflicting values — two memories give different values for the same attribute of the same subject.
• Scope mismatch — the memories are about a different target than the one asked about (right attribute, wrong entity).
• Insufficient — nothing recalled actually contains the asked-for value.

If none hold, answer directly from the recalled memory.

Step 3 — Re-query (the self-correction)

For an ambiguous or scope-mismatched result, run a narrower recall naming the specific entity and attribute:

simba memory recall "<specific entity> <attribute>"

Repeat once or twice with sharper terms. Broaden only if nothing returns:

simba memory recall --limit 8 "<broader phrasing>"

Step 4 — Resolve conflicts by recency

When two memories conflict, prefer the fresher one:

• Recalled context tags the most recently created memory with recency="newest" and shows each memory's created date — prefer it.
• For knowledge-graph facts, the currently-valid edge is the one with no valid_to; check with simba db facts (it prints occurred: event dates).

Step 5 — Answer or ask

• Clear winner → answer, citing the specific memory.
• Still ambiguous after re-query → ask the user to disambiguate ("Do you mean the staging or prod key?"). Do not guess.
• Not in memory → say so plainly ("I don't have that in memory"). Never fabricate a value to fill the gap.

Verification before you answer

• The value you're about to give is attached to the exact entity asked about.
• You did not mix attributes/values across different entities.
• Partial or low-confidence answers are disclosed as such.