What does your agent forget between conversations?

A question for anyone running agents in a day job rather than in a demo.

Mine used to lose things between conversations, and I only noticed because I kept retyping them. None of it was hard for the model. It was just gone by the next morning.

What I ended up doing was making the memory a file rather than a habit — a title and a body, on disk, that the job opens already holding, and every other job can read.

The harder half is what happens when two of those notes disagree. Facts go in a base of their own, one plain sentence each. A new one is compared by meaning against everything already in there, and every near pair stops and asks for a ruling — both true, merge them, one wins, or drop the new one. Nothing commits while a pair is open.

What's the thing you re-supply most? I'm collecting them, because I want to know whether they're the same for everyone or just mine.