What's Really Happening When You Chat

If you've used one of these tools for any real stretch, you've smacked into some genuinely strange walls. It forgets what you told it yesterday. You paste in ten pages to be helpful, and somehow it gets worse. You correct a mistake, and an hour later it makes the exact same one.

These feel random, a little maddening even, like the tool is just flaky. It isn't. There's one tidy idea sitting underneath every one of those moments, and once you see it, the chatbot stops being mysterious. Here it is, plain: the AI only knows what's in front of it right now. That's the whole lesson. Everything below is just that idea wearing different coats.

It re-reads — it doesn't remember

Start with the thing that surprises people most. When you hit send, the AI doesn't recall your conversation the way a friend remembers a chat you had. Every single time, it re-reads the entire conversation from the top (every message, yours and its own) and only then writes the next reply.

That's the gotcha: it isn't remembering, it's re-reading. (Which fits what we learned two lessons back — it has no memory of its own; each turn, it's generating fresh from whatever text is in front of it.) And it leads to one rule that explains almost everything else:

The desk it works on

Picture a desk. A fixed-size desk, and the whole conversation is laid out on it — every message, stacked up where the AI can see it. Anything on the desk, it can read and use. That's its working space, and it has a name people throw around: the context, just "what's currently in view."

How big is the desk? It's measured in tokens: think of those as small chunks of text, roughly what you're "spending" as you chat (not quite the same as words, but close enough for now). You don't need the exact definition; just hold onto the picture of a desk with a fixed amount of room.

And here's the part that resolves a real frustration — the desk has an edge. Pile on enough and the earliest pages slide off to make room. So when a long chat starts drifting, or seems to "forget" what you said way up at the top, that's not a glitch. The top of the conversation has quietly slid off the edge of the desk.

A new chat is a clean desk

Open a brand-new chat and the desk is wiped clean. Nothing from before is on it.

That sounds like a limitation, but flip it around — it's one of your most useful tools. When a conversation goes sideways, gets cluttered, or wanders somewhere unhelpful, you don't have to wrestle it back on track. You just start a fresh chat and begin again on a clean desk.

Two chats never compare notes

Here's the flip side of a clean desk. Chat A has no idea Chat B ever happened. Different desk, completely sealed off.

This clears up the single most common surprise people hit: "but I told it yesterday…" You did — in a different chat, on a different desk that this one simply can't see. Nothing carried over, because nothing ever travels from one desk to another on its own.

Memory: the one deliberate exception

Now, you may have heard that some AI tools can "remember" you across chats — and that's real. But it's worth seeing exactly how it differs from the desk.

The desk (the context) is automatic and temporary: it fills up as you chat and resets every new conversation, with no effort from you. Memory is the opposite: a deliberate, opt-in feature that some tools offer, where the app keeps a small, lasting note about you and quietly sets it on the desk at the start of each chat. It's usually something you can see and edit.

Keep the distinction simple: that lasting note isn't the model remembering — it's the app taking notes for you and laying them back on the desk. Whether a given tool offers it, and what it's called, is something we'll save for the later lesson on tools and settings.

Correcting it doesn't teach it for good

This is the big one — and the most common misconception in the whole territory. When you correct the AI and it says "you're right, my mistake," it feels like you just taught it something. You didn't, not permanently.

Your correction helps inside that one conversation — because now it's sitting right there on the desk, and the AI re-reads it on the next turn. But it does not update the AI for next time. (Remember: it isn't learning from your chats. It was trained once, then frozen.) So when you correct something and an hour later it's wrong again, here's why: you're either on a different desk, or you're far enough down the same one that your correction has slid off the edge.

It all comes back to the desk

So every wall you've hit is the same idea in a different coat. It forgets you: different desk. The long paste made it worse: crowded desk. Your correction didn't stick: it slid off the desk, or it was on another one. It only knows what's on the desk right now. Hold that single picture and the chatbot stops being unpredictable.

And here's the practical payoff. Since it only works with what's on the desk, and improvises from there, what you put on the desk decides what you get back. Which means the single biggest upgrade to your results isn't a better tool; it's learning to ask well. That one skill, how to prompt so it nails the answer the first time, is exactly where we head next.