The Context Window

It is late afternoon, and Tessa has been going back and forth with the chat box for an hour — refining a brochure, trying angles, pasting drafts. Then something odd happens: the model's latest suggestion cheerfully contradicts an instruction she gave near the start. "Keep every mention of pricing out of the text," she had written, and here is pricing, back in the text. Nothing broke. She has just met the model's most important limit, and it deserves the plainest possible explanation.

The model has exactly one working memory, called the context window: the amount of text it can consider at once, measured — as everything is — in tokens. Everything that should influence the next answer has to fit inside it: your messages, the model's own replies, anything you pasted. It is big — modern windows hold a book's worth of text — but it is never infinite. Picture the whiteboard in a meeting room: everything the meeting has decided must be written on the board, and when the board fills up, someone erases the oldest corner to make room. The model answers from the board. Never from the meeting's memory — there is no meeting's memory.

What Happens When It Fills

An hour of back-and-forth adds up. Every message and reply lands in the window, and once the conversation outgrows it, the oldest turns stop fitting — Tessa's pricing instruction among them. The model then answers using only what remains in view. From the outside this reads as forgetting; mechanically, it is simpler and stranger: the instruction was not forgotten, it was no longer shown. The model followed everything it could see, perfectly. What it could see no longer included her rule.

This explains a family of frustrations that every regular user has met. Long chats drift off their brief. Constraints quietly lapse. A carefully established persona goes flat. The model contradicts "what you told it" — except you told it something that has since slid off the whiteboard. Nothing is malfunctioning, and — this is the useful part — nothing is random about it. Overflow has symptoms you can now recognize on sight.

Every Chat Starts Empty

There is a second half to the picture. Open a new conversation and the window starts blank: the model carries nothing over from yesterday's chat, or from the chat sitting in the next browser tab. Whatever it seemed to know about Tessa yesterday, it does not know now. (Why the chat box nonetheless shows her old conversations, and how some products give a real impression of remembering you, is a wrapper's trick — Chapter 7 reveals it, and it is a good one.)

For now, take the practical fact: a conversation is a self-contained world. Everything the model should know in this world, you put in this world.

Living with the Window

Three habits turn this limit from a trap into a tool, and they cost nothing. Start fresh for a new task. A long, wandering chat drags its whole history through the window on every turn; a clean chat starts sharp. Restate what matters. If an instruction must hold, repeat it in your latest message rather than trusting the one from an hour ago — professionals do this constantly, without embarrassment. Put the important thing near the end. Recent text is always safely in view; ancient text is what falls off.

And when Tessa meets a document too big for the window — the 120-page supplier contract is waiting in Chapter 4 — the answer will not be to push harder. It will be craft: choosing what goes onto the whiteboard, because choosing is the skill the window forces, and it turns out to be the skill that separates good model use from bad far more than clever phrasing ever does.