One thing that worries me with “learn from anything” systems is that the memory layer effectively becomes part of the attack surface. We’ve seen cases where stale or malicious context persisted far longer than expected and influenced future agent behavior.