When should an AI agent be allowed to remember?

Agent memory can be useful, but a confident guess from yesterday can turn into a trusted fact today. In our QA process we decided that a remembered selector or repair should only be returned when it points to concrete evidence: a verified run, a passing test, or a repair that later passes. Recency and contradictions can lower its confidence. No evidence, no memory.

If you’re building agents, what criteria make a memory trustworthy enough for automatic reuse?