The Illusion of Conversation

Tell a chatbot your name and, twenty minutes later, it uses it. It refers back to the thing you said earlier, builds on its own suggestions, keeps a thread going. After Part 2, that should bother you. The file is frozen. The engine only finishes text. So who, in this picture, is doing the remembering?

Nobody is. This part is about the single most important trick in the whole stack, and it is so low-tech that people refuse to believe it at first.

Nothing is remembered, everything is re-read

The model has no memory between messages. None. Not short-term memory, not a summary, not a fading impression. The moment it finishes writing a reply, everything is gone.

So the software behind the chat window does the only thing that can work. Every time you press send, it pastes together the entire conversation so far, every message of yours and every reply of its own, into one long text, and hands the whole thing to the engine. The model re-reads the conversation from the very beginning, every single turn, and its "reply" is a continuation of that ever-growing transcript. Then it forgets again, and the cycle repeats on your next message.

Said plainly: every time you message an AI, it reads your whole conversation from the start, because it has already forgotten all of it. Its memory is the transcript. When the AI "remembers" that your name is Sam, what actually happened is that "My name's Sam" is still sitting in the text being re-sent, and a machine that continues text will naturally continue it in a way that uses the name. The illusion is airtight only because the re-reading is fast and completely invisible.

Between your messages, nobody is home

A companion misconception is worth clearing at the same time: the model is not sitting there thinking about your problem while you type. Between messages, the model is not running at all. The engine produced the final token of its last reply and stopped, the way a calculator stops after showing a result. There is no process in the background musing about you, wondering where you went, or preparing its next point. When you press send, the whole apparatus spins up, re-reads, continues, and goes silent again.

Watch the trick happen below. The left side is the chat as you would see it. The right side is what the engine is actually handed on each turn.

Step through it and watch the right-hand side. The thing to notice is not what gets added, it is that the whole stack is present every time. Turn five hands the engine turns one to four all over again. Then press fast-forward and watch what a long conversation does to the desk.

The size of the desk

That desk has a name: the context window, the maximum amount of text a model can be handed in one go. Depending on the product, it ranges from around 8 thousand tokens to over a million, and the number is one of the few honest specifications on any AI datasheet. A conversation that outgrows the window has a problem with no clean solution. The oldest material falls off, or the app quietly summarises earlier turns into a shorter form, and either way the model is not being coy when it misremembers what you said an hour ago. For the model, that text no longer exists.

The desk analogy breaks in one way worth flagging: a real desk gets gradually messier, while a context window is a hard edge. Text either fits on it or has fallen off it, and the falling is abrupt.

The transcript trick also has a price tag that surprises people. Because every turn re-sends everything, the hundredth message of a conversation hands the model vastly more text than the first one did, and on paid services you are billed for all of it, every time. Long chats do not cost more in total only, they cost more per message as they go. Hold that thought for the final part.

"But my app does remember me"

Some AI apps advertise memory across conversations, and now you can see exactly what that feature is. The app keeps a notes file about you ("Name: Sam. Prefers short answers.") and quietly pastes those notes into each new conversation's text. Useful, and a reasonable thing to build. But it is the same transcript trick wearing a different coat, notes travelling in with the text, not weights changing. The file on the disk has still never heard of you.

Our running example has now gained some luggage. "What's the weather in Paris?" will never reach the model alone. It arrives stapled to the entire conversation before it, and possibly to a notes file too. Which raises the question Part 4 answers: stapled together how, exactly? It turns out the pasting has strict rules, the rules differ for every model family, and when they go wrong, the weirdness you see is one of the great unexplained experiences of using AI.

Next: The Script Rewrite: How a Chat Becomes One Long Text: roles, chat templates, and the hidden briefing document that gives every chatbot its personality.