Part 1 of 4 in The Brilliant Stranger, a field guide to LLMs, explained like you run a business.
“Why doesn’t it remember what I told it?”
You explain your pricing to the AI on Monday. On Tuesday you ask a follow-up and it has no idea what you’re talking about. Or you ask about a deal that closed this morning and it confidently makes something up. That looks like a defect. It isn’t. It’s the most important thing to understand about how these systems work, and almost nobody explains it.
That’s what this series is for: one analogy per post, for the people who make decisions about AI. We start with the question almost everyone asks first.
The wrong picture: the chat window is the AI
Most people assume the thing they’re typing into is the AI. One box, one brain, and like any brain, it should remember what you told it. Under that picture, forgetting is a bug.
The better picture: an expert in a room with no windows
The model itself is a brilliant expert sitting in a sealed room. No windows, no phone, no clock, no filing cabinet. They know everything they read before the door closed, and nothing since. They can’t look anything up or check the date. They have no idea what company you work for.
The only way anything reaches them is a note slid under the door.
And every conversation is a fresh visit. When you come back tomorrow, the expert has no memory of you. Nothing from yesterday carries over unless someone writes it down and slides it under again. That’s what people mean when they call a model stateless: it meets you for the first time, every time.
It’s true inside a single conversation, too. Every time you send a message, the whole conversation so far gets slid under the door again with it. That’s why a long chat gets slower, costs more, and starts to lose the thread.
The assistant outside the door
So how does it ever know anything about your business? The product you actually use is not the expert. It’s the assistant standing outside the door. The industry calls this the harness.
When you ask a question, the assistant decides what to slide under the door: your message, maybe a few pages from your documents, maybe notes from your last conversation. If the expert needs something it doesn’t have, like this quarter’s numbers or today’s date, it asks, and the assistant goes and gets it.
When an AI tool seems to remember you, this is what’s happening. The assistant kept notes and slid them back under the door. The expert didn’t remember a thing.
Two limits the door puts on everything
The door has a size. Only so many pages fit through at once. That limit is the context window, and it’s why you can’t slide your entire company under the door. Every page costs money, every time, and past a point the expert starts skimming.
And the door only opens from the outside. The expert can’t reach out on their own. They can ask for something, but the assistant decides what actually comes through. So every question about “what does the AI know” is really a question about what the assistant was built to hand it, and someone designed that.
What this means for your business
Stop judging the model on its own. When a tool gives a bad answer about your business, the expert is rarely the problem. Ask what was slid under the door and what wasn’t.
Treat memory as something you build. Nothing persists on its own. If you want it to know where a deal stands or what was decided on last week’s call, something outside the door has to store that and hand it back.
Decide what the assistant can go get. Every errand it can run is a door into your systems. That’s where the power is, and where the permissions conversation belongs.
Next time: if the expert can’t remember, why not teach it permanently? That’s where the guitar comes in.
Trying to figure out what your AI should actually be seeing? Reach out.