What is Context Window? — one analogy

1The analogy

A clothes drying rack has a fixed amount of room — a set number of bars, a set length. However much laundry there is, none of it can hang past what the rack holds.

There's only one way to hang something new: take down something already hanging. Take it down and fold it into a drawer, and it disappears from the rack. Only what's currently hanging is visible at a glance.

What each item takes up differs too. A sock needs barely any room; a big towel eats up a whole section on its own. It's the bulk that decides the space, not the count of items hanging there. The range a model can take in and look at at once works the very same way.

2In detail

Space gets measured in pieces

The way a sock and a towel take up different amounts of room on the rack, length alone doesn't say how much room a piece of writing needs. A model handles text by slicing it into pieces, and the piece count is exactly what eats up the space.

Carrying the same meaning, English tends to use fewer pieces than a language like Korean. So a document that looks like about the same length can take up far more room in one language than another. A document full of tables or lists, or one packed with numbers, eats up more pieces than it looks like it should too.

That's why a limit gets shown as something like "a few hundred thousand pieces" rather than a character count — character count simply can't pin it down precisely.

One space, shared by several things

The rack doesn't only hold today's laundry. Everything spread out in front of a model — the instructions a service inserted ahead of time, the conversation so far, this question, and any pasted document — all pile onto that same space together.

Room even has to be set aside ahead of time for the answer still to come. Fill the space entirely with the question and there's nowhere left for the answer to land. A long document pasted in whole, followed by an answer that suddenly comes back short or cuts off mid-sentence — this is usually why.

Overflow, and the oldest gets taken down first

Run out of room as a conversation grows longer and the oldest parts get pushed out first. Say "don't use formal language" at the very start, and see formal language creep back in much later — that request has already been taken down off the rack.

The model hasn't forgotten. It's simply reached a state where it can no longer see it at all. There's no memory being searched — every single time, only whatever is currently hanging on the rack gets read from scratch.

How the pushing-out happens differs by service. Some just cut off the earlier part; others compress the earlier conversation into a short summary to save room. A summary keeps the gist but loses the fine wording.

Is a bigger space only ever a good thing?

Swap in a bigger rack and more fits at once. It also takes up more of the room. A context window is the same. A bigger one lets more get pasted in, but reading through all of that from scratch every single time makes the answer slower and the cost climb along with it.

There's a middle-blur problem too. Hang a great deal on the rack and there's a tendency to keep a close eye on the beginning and the end while missing whatever's somewhere in the middle. A bigger space doesn't mean everything inside it gets looked at with equal care.

How to use the space wisely

The surest approach is hanging up only what's actually needed. Rather than shoving in a thick document whole, finding and pasting in just the relevant part saves room and sharpens the answer too.

For a long conversation, it helps to summarize partway through. Write the key points down in a few lines and start a new conversation, and it works like folding the taken-down laundry into a drawer. Any condition that really matters is safer repeated where it counts than left only at the very start of the conversation.

3More precisely

A context window is the cap on the number of tokens a model can handle in a single pass. Input and output are often counted together against that one shared cap, though some services set separate limits for each side instead. This cap is set by the model's structure and how it was trained — it isn't a setting a user can adjust, and it stays fixed no matter how a request gets phrased. A large cap also doesn't guarantee quality holds up when it's actually filled that far; the length a model actually handles well in practice is often noticeably shorter than the number advertised on a spec sheet.

The analogy breaks down somewhere. Laundry left hanging on a rack stays put if nobody touches it, but a model reads everything back in from scratch with every single exchange. Nothing stays hung in place — it's laid out fresh each time. And a rack dries anything hung on it equally as long as there's room, but the more that's packed into the space, the thinner the attention the model spends on each part of it.

4Try it yourself

Tiktokenizer (see the split) ailearn.space Count how many pieces a piece of writing breaks into to get a feel for how much room it would actually take up

5Common misconceptions

It's easy to think a context window is a model's memory, but actually it's just the range read back in each time — anything outside it might as well never have existed.
It's easy to think counting characters tells you how much room is left, but actually it's counted in pieces, so it swings a lot depending on the words chosen.
It's easy to think a model with a big window sees everything inside it equally well, but actually content sitting somewhere in the middle gets missed fairly often.

7One-line summary

In shortA context window is the size of the space a model can lay out and look at at once, and whatever falls outside that space is invisible to it from the start.

Did this analogy help?

Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02