Every memory carries a confidence built from how it was captured, how often other memories have confirmed it, and how fresh it is. Tiers run from VerifiedFact (0.90+) and HighConfidence (0.70+) down through Moderate, Low and Stale.
What your assistant sees
On every search hit:
assertable — true means the assistant may state it as fact; false means "a lead to verify with you".
query_fit — how well this memory answers this question. A trusted memory that barely fits isn't the answer.
memory_confidence — the tier of the stored memory itself, which is the same whatever you ask.
Assistants following the JustOnce protocol only assert what is assertable, cite what shaped their answer, and say "nothing stored" rather than inventing.
In the app
A memory's Details card shows its confidence. Memories below 0.70 — the ones assistants treat as hints — collect under Needs review in the filter menu on Home, so you can confirm, edit or delete them.
Conflicts
When two memories disagree — one says the rego is due in March, another says April — the hit is marked conflicted. The assistant asks get_as_of for the arbitrated value (the most recent, best-supported one); ChatGPT and Codex aren't offered get_as_of. Either way, you can resolve it for good by correcting or deleting one of them in the app.
Making a fact stronger
Confirm it: say it again to any assistant, or edit the memory so the details are clear. A new memory that repeats a detail reinforces the earlier one, and reinforcement raises confidence.