The Context Window Is a Desk, Not a Memory

Where does the model keep what you told it twenty minutes ago? You explained your project, it answered with your constraints in mind, so the knowledge must be sitting somewhere. Look for it. There is no drawer. There is no filing cabinet in the back. There is only the desk, and the desk gets repacked from scratch every single turn.

The desk

Between your messages, the model holds nothing. It does not keep your conversation warm in some buffer, does not mull over what you said, does not wait for you. It is not even idle in any meaningful sense. The model only exists for you during a call: input arrives, a reply comes out, and everything is gone.

So how does a conversation work at all? Every time you send a message, the product you are using gathers the entire conversation, the system instructions, every message you wrote, every reply it gave, any documents you attached, plus your new message, and sends all of it again. The model reads the whole thing as one input, top to bottom, and produces a reply as if it had been present all along. Then it forgets everything. Next turn, the same ritual.

The context window is the size of that working surface, and it is measured in tokens, never in messages or minutes. Everything shares it. Instructions, history, documents, your question, all on one desk. Nothing the model "knows" about this conversation lives anywhere else.

Why long chats cost more

Once you see the re-send, the bill explains itself.

At message 5, asking a short question sends a short conversation. At message 50, asking the same short question sends fifty messages of history along with it, because the model needs all of it to be the assistant you have been talking to. You pay for the desk, not for the question. The same "what about option two?" costs a few tokens of new text and a whole conversation of context.

This also means your instructions pay rent. A system prompt, a style guide, an attached document: they ride along on every single turn, priced every single turn. A heavy prompt is not a one-time setup cost. It is a recurring charge for as long as the conversation lives.

Why the model "forgot"

The desk has edges. When the conversation outgrows the window, something has to go, because all of it no longer fits.

What goes depends on the product. Some drop the oldest messages. Some summarize the early conversation and keep the summary instead. Both are losses. A summary is not the thing itself, it is a compressed account of it, and the detail you need later may be exactly what compression smoothed away. That is why the model loses the file name you mentioned an hour ago, or re-asks a question you already answered. The conversation got long, the original fell off the desk, and what remains is a paraphrase.

The forgetting is mechanical, not moody. It does not fade gradually the way human memory does, and it does not pick what to lose based on importance to you. It happens exactly when the conversation outgrows the window, which means you can predict it, and you can act before it.

What "memory" features really are

Some products remember you across sessions. You start a fresh chat and it knows your name, your stack, your preferences. Given everything above, that should look impossible. The model that greets you today holds nothing from yesterday.

Here is the trick: the model did not remember. Something outside it did. The product wrote notes about you, stored them in an ordinary database, and quietly places them on the desk at the start of every conversation. The model reads them along with everything else and replies as if it knew you. It is real and it is useful. It is also just context, prepared by the product instead of by you. The desk is still the only thing the model ever sees. The continuity is staged, and the stagehand is software you do not see.

The fresh session is a technique

This reframes a thing people treat as failure. When a long chat starts dragging, going in circles, losing track, the instinct is to push through, because starting over feels like losing everything.

But you now know what the model actually has: a crowded desk where the important things may already be paraphrased or gone. A fresh session with the essentials restated puts exactly the right material on a clean desk and nothing else. Three sentences naming the goal, the constraints, and the current state often beat fifty messages of drift.

You are the memory in this arrangement. The model brings the reasoning. You decide what deserves a place on the desk. Curating that surface, what gets restated, what gets attached, what gets left behind, is the actual skill of working with these tools.

Close

The model arrives at every call the same: frozen, blank, and ready to read. Whatever continuity you experience was placed on the desk this turn, by you or by the product on your behalf.

The loop runs on tokens, and the desk is where all of them have to fit. There is no drawer. There never was. Pack the desk well, and the model will look like it remembers everything that matters.