Your agent gave a wrong answer. One command finds the memory that caused it.

Description and use

The mnemosy bisect command in Mnemosyne 3.0 quickly identifies the record causing a wrong agent answer, while new correction and poisoning‑gate features maintain memory integrity.

Problem

An AI agent’s memory contains stale or incorrect records, causing wrong answers and making it hard to identify the culprit record.

Outcome

On a 2,000‑record synthetic store with one seeded false record, mnemosy bisect achieved precision 1.00 in ~1 s. The poisoning gate prevented 0/167 escapes and 0/60 false quarantines. Dependency correction achieved recall 96.05 % and precision 98.06 % with zero stale procedures when the premise guard was used.

Workflow

• 1. Run mnemosy bisect --db memory.sqlite --workspace demo --agent a --json '{"id":"key-rotation","query":"deploy key rotates","k":1,"gold":["e233b376…"]}' to locate the culprit record.
• 2. Use correct() to retire dependents and propose inferred dependencies, reviewed by a human.
• 3. Verify that the poisoning gate blocks untrusted instructions from being promoted.
• 4. Undo unwanted consolidation with the provided undo command.
• 5. Re‑run the agent to confirm the correct answer returns.

Evidence

mnemosy bisect names the record behind a wrong answer. On a 2,000-record synthetic store with one seeded false record, it found exactly that record (precision 1.00) in about a second.

Source

The number: precision 1.00 in about 1 s on 2,000 records.

Source

The number: recall 96.05 %, precision 98.06 %, 0 stale procedures with the premise guard.

Source

The number: 0/167 escapes, 0/60 false quarantines.

Source