“Keep this scene dark” should not become a permanent preference for every scene I make.
That sounds obvious, but persistent memory can turn a temporary instruction into a lasting assumption. Later, the assistant appears to know me while repeatedly making a choice I no longer want.
I’d separate explicit preferences from inferred ones, and project instructions from account-wide settings. Each saved item needs a source and a scope. Temporary instructions should expire or come back for confirmation.
The interface can be simple: remember this for the project, remember this generally, or use it only today. I also want to see and correct the saved items. Editing a memory should be easier than arguing with its consequences in another conversation.
Deletion is the fiddly engineering part. Removing the original note doesn’t help if a summary, embedding or cached prompt still carries it forward. Follow the derived copies and invalidate them too. “Forgotten” should mean more than hidden from the memory screen.