The window, in plain terms

The most common wrong belief about these tools is that they remember you. They do not. Everything a model can use when it answers is laid out in front of it at that moment, and anything not laid out in front of it does not exist.

The industry calls that space the context window. The word window is unhelpful — nothing is being looked through. A desk is closer.

The desk

Picture a desk of a fixed size. Before a single word of an answer is written, everything the model is allowed to use gets laid out on it:

• the instruction you just typed
• everything said so far in this conversation — your messages and its own replies alike
• anything you attached or pasted
• any standing instruction your settings, your workspace or your organisation applies to every conversation

That is the whole of it. There is no filing cabinet behind the desk. Last Tuesday’s conversation is not there. The report you attached in a different thread is not there. Your pricing policy and the thing you explained twice last month are not there either, unless something put them on the desk this time.

The second property matters as much as the first. The desk is laid out again from scratch every turn. When you press send, the entire contents are read from the beginning: the original brief, the eleven exchanges since, the attachment, your new sentence. The model does not recall its previous answer. It re-reads it.

That single mechanism explains three things people find odd. Long conversations get slower — and on metered plans cost more per question — even as your questions get shorter, because the pile being re-read keeps growing. A tool can quote something from forty turns ago word for word, because it is still on the desk. And it can lose that same thing entirely a few turns later, because it is not.

The edge

The desk holds only so much paper. What happens when a conversation exceeds it is a property of the tool rather than of the model, and there are three behaviours in circulation.

1. It refuses. You are told the conversation is too long. This is the cleanest outcome, because you know.
2. The oldest material is dropped to make room for the newest. The conversation carries on, apparently normally. Nothing announces the loss.
3. Older material is compressed into a summary of itself. Better than dropping it, and still lossy — a summary keeps the gist and sheds the specifics, and the specifics are usually your constraints.

The second and third are the dangerous ones, and it is worth being blunt about what they mean. What falls off the edge is gone. It is not archived, not retrievable, not in there somewhere. From the model’s side it never happened.

In practice it looks like this. A commercial director opens a conversation with a careful brief: the pricing floor, the two competitors not to be named, the structure the board expects. Then forty turns of drafting against a long tender. Somewhere after turn thirty the drafts start quoting below the floor. Nothing broke, no error appeared, nobody changed the request. The brief left the desk, and every answer since has been written by something that never saw it.

What is sharing the desk with you

Two things routinely take more room than people expect.

Attachments. Where your tool lays a file out whole rather than storing it and retrieving pieces from it, the room it takes is room the conversation does not have — and attaching four reports to a discussion that then runs for an hour is the quickest route to the edge.

Its own answers. Every reply stays on the desk and is re-read with everything else. Asking for full drafts where an outline would do fills the space faster than people expect, because you are paying for the request and the output both.

What changes on Monday

Three habits follow directly, and none of them require you to know anything about how any of this is built.

Restate what binds you. If a rule matters and you stated it half an hour ago, state it again. Not as a rebuke — as a delivery.

Assume nothing carries between conversations. A rule that must apply every time belongs in a document you paste or a standing instruction your tool applies automatically. It does not live in the tool’s head, because there is no head.

Watch what you are spending the space on. A conversation carrying three attachments and forty turns of drafting has very little room left for the thing you actually care about — and it will not tell you so.

Checkpoint

A long tender conversation starts producing drafts that breach a pricing rule you stated at the outset. Nothing else changed. What is the most likely explanation?

The model is not capable enough for commercial work of this kindThe rule was ambiguous, so it settled on a different readingThe rule is no longer in front of it — the desk filled up