This is the core problem. The agent writes its own memory while working, so it has blind spots about what matters. I've had sessions where it carefully noted one thing but missed a bigger mistake in the same conversation — it can't see its own gaps.

A second pass over the transcript afterward catches what the agent missed. Doesn't need the agent to notice anything. Just reads the conversation cold.

The two approaches have completely different failure modes, which is why you need both. What nobody's built yet is the loop where the second pass feeds back into the memory for the next session.