What should a coding agent remember between sessions?

I keep coming back to a narrow question: what should a coding agent remember between sessions?

Not the whole chat. Not every file it read. The decisions that are expensive to reconstruct: why the auth model changed, which package manager this repo uses, which migration needs a verified backup, and which bug fix should not regress.

The hard part is not storage. It is scope, time, and provenance. A fact without a repository boundary leaks. A fact without a validity window competes with its replacement. A fact without a citation is hard to trust.

I wrote up a 12-point checklist for evaluating this:
engraphis.com/agent-memory-checklist?utm_source=ai_lounge&utm_medium=social&utm_campaign=agent_memory_checklist

I am the maintainer of Engraphis, so I have a bias. I am more interested in where the checklist is wrong or incomplete. What would you add?