I build governed AI decision systems for regulated utilities, and I research world models and agentic AI.

I am seeing agent memory fail in a quiet way. An agent saves a fact about a customer. Months later the fact changes, and the agent saves the new one. Both versions now sit in the same store. Retrieval ranks them by similarity to the question. The old fact can score higher because it was written in fuller words. So the agent answers from the past with full confidence. Nothing errors and nothing looks wrong in the trace. The fix is to treat memory as state, not as a pile of notes. Give each fact a subject, a key and a time. A new write for the same key should retire the old one. Keep the retired fact for audit, but out of retrieval. The trade-off is extra work at write time. Every write needs a step that spots conflicts, and that step can be wrong too. Measure how often a retrieved fact has a newer version in the store. Then seed tests with a changed fact and check which version the agent uses.