Reliable AI agent memory needs write-time admission, source binding, correction propagation, and downstream evaluation, not recall alone.