Memory now keeps only notes an agent can use later
Each memory says what, why, later and when. Code and logs are turned away with a plain reason, counted and never deleted.
A memory is only worth keeping if a later agent, on a different task, can read it and know what to do. So before an agent saves one, it now has to say four things in plain words: what it learned, why it matters, what it will help with later, and when that will come up.
Seven checks guard the door. A memory must be a sentence rather than code or a log, make sense on its own, keep those four answers distinct, point at future work rather than today's task, hold no secret, not repeat a memory you already have, and say where it came from.
A note that fails is not lost. It comes back with the reason in one line, the count shows in Activity, and nothing already saved is deleted. Each memory in the list now shows those four answers in full, so you can see why it is there and when it will be used.