Reveal on the spot: before anything else, tell the person everything you know about them right now. Who they are, what they do, how they use AI, anything from past conversations. Then say where each piece came from: a memory the product saved, a file you were loaded with, or this conversation. Then ask them to correct you. One round. What comes back is the lesson.
Right. They said it, and it got kept. Nothing mysterious.
Half right. It was true when saved and isn't now, or the detail got flattened into a summary. This is the worst one because it doesn't announce itself.
Missing. They said it, but in a conversation that ended and nothing carried over.
Wrong. The AI inferred it, or filled a gap to be helpful.
What to know: memory is not magic. It is a file sitting somewhere that the AI saved, holding context about you. When it starts a conversation, it reads that file. That is all "remembering" is.
When a product keeps memory for you automatically, the AI decides what is worth keeping and writes it down without asking. That solves re-explaining yourself, a bit. It does not solve the real problem, which is getting the right detail in. With automatic memory, you don't choose what gets kept, how it gets summarised, or when it goes stale. The four cases above are what that looks like from the outside.
Author's practice: own where the knowledge lives, and manage it. The context the AI has about you should be in files you can read, edit, and point at, not ones the product writes on your behalf. But one big file holding everything you know is just a different way of polluting the window. The AI would read all of it, every time, for a question that needed one line. So memory is kept the way the standing file is kept: the least that every conversation needs, and pointers to the rest, so the AI fetches a detail only when the work calls for it. The standing file, in Intermediate, is where that starts.
Unlocks: knowing exactly what the AI thinks it knows about you, and being the one who decides.