What should a coding agent remember between sessions?

I keep seeing the same failure: the agent can solve a task in one session, then starts the next session without the decisions that shaped the code. I built Engraphis to make memory scoped by repo and session and inspectable. Before launch tomorrow, I am curious what people here think is actually worth remembering: architecture decisions, bug fixes, team conventions, failed approaches, or something else? And what should the agent deliberately forget?