How do you find out afterwards whether your agent's memory was right?

I'm working on long-term memory for agents and trying to figure out if a problem we keep seeing in tests is real for people running agents in practice, before building anything for it. In our tests, a planted copy of a real memory with the conclusion flipped wiped out the whole benefit of having memory. The only thing that stopped it was letting a memory drive a decision only if its outcome was later confirmed from outside the agent (a test passing, a ticket closed, a person signing off). When the agent graded its own memories, it happily confirmed the planted one too. So, for anyone running agents with long-term memory: When your agent acts on something it remembered, do you ever learn later whether that memory was right? What's the signal? Has a stale or wrong memory ever caused a bad action you had to trace back? How did you find which one it was? Do you pay for anything today to keep agent behaviour safe or auditable? "We never get that signal" is a genuinely useful answer.