Glossary
Context window
A context window is the maximum amount of text a model can take into account for one reply, covering the conversation so far, anything attached and the answer being written, beyond which the oldest material stops counting.
In plain terms
Think of a desk with a fixed amount of space. Everything the assistant is working from has to be on that desk at once: the question, the files, what you both said earlier. When the desk fills up, something has to come off, and whatever comes off is simply gone as far as this reply is concerned. A bigger desk does not make the assistant cleverer, only better supplied.
Why it matters
Almost every complaint about assistants losing the plot traces back to this one limit. The long conversation that starts contradicting itself, the attached contract whose later clauses seem to be ignored, the instruction given twenty messages ago that no longer holds: none of those are the model being careless, they are material that has fallen off the desk. Knowing the limit exists changes how you work. You stop expecting continuity that was never on offer, you start new conversations deliberately rather than letting one sprawl, and you attach the ten relevant pages rather than the two hundred page document.
How it works
The window is measured in tokens rather than words or pages, and every part of the request draws on the same allowance. Any standing instructions the product sets, the whole visible conversation, every file you have attached and the reply being generated all compete for one budget. A window advertised as generous can still be tight in practice if a long system instruction and three attachments are consuming most of it before you type.
When the total would exceed the limit, something must give, and what happens next is a product decision rather than a property of the model. Some tools refuse the request outright. Some silently drop the oldest turns. Some summarise the earlier conversation and carry the summary forward instead of the original wording. The third is the most useful and the most deceptive, because the assistant continues confidently from a compressed version of what was said.
Capacity is not the same as attention. Material placed in the middle of a very long input is more easily under-weighted than material at the start or the end, an effect consistent enough that it has a name in the research literature. This is why extracting the relevant section often beats attaching the whole file, even when the whole file fits comfortably inside the limit.
The window is also the reason long conversations cost more per message. The history is resent on every turn so the model can follow the thread, so the input grows steadily even when your questions do not. Cost and the risk of overflow climb together, which is why starting fresh when the subject changes helps on both counts at once.
Nothing persists between conversations by virtue of the window alone. Once a conversation ends, its contents are not retained by the model. Any continuity you notice across sessions is a separate feature storing notes and re-supplying them, which is itself consuming part of the same budget when it does.
Retrieval is the standard way round the ceiling. Rather than attaching an entire library, a system searches it for the passages relevant to the question and inserts only those. The window stops being a limit on how much material you can own and becomes a limit on how much you need in front of you at once, which is a far easier constraint to live with. That is the whole reason the pattern exists.
Advertised sizes and usable sizes are not always the same number. Some products reserve part of the window for their own instructions, some cap attachments below the model's ceiling, and quality can taper before the hard limit is reached. Treat the headline figure as the ceiling rather than the working space, and test with your own longest realistic document rather than trusting the specification.
What a window holds, in units you already know
Seen in the wild
Attach a long report to an assistant built for large volumes of source material and ask about a detail in the final section, which is the case where a big window earns its keep.
ClaudeWatch a long chat begin repeating a question you already answered, which is the earliest turns having dropped off rather than the assistant being careless.
ChatGPTRun a published model on your own machine and find the window is smaller than a hosted service offers, because it has to fit the memory you actually have.
Ollama
Common misconceptions
People assume
A bigger context window means the model understands more.
In fact
It means the model can be shown more. Comprehension, accuracy and judgement are unchanged by the size of the window, and material buried in a very long input is more easily under-weighted than the same material presented on its own. Capacity and attention are different things.
People assume
Once the window is full the conversation stops working.
In fact
Usually it keeps going, which is the problem. Most products quietly drop or summarise the oldest material and carry on, so the assistant answers confidently from a version of the conversation that is missing pieces. The failure is silent rather than loud, so nothing tells you it happened.
People assume
Anything I said earlier in the conversation is remembered.
In fact
Only while it is still inside the window. Earlier turns are resent with each new message, and once they no longer fit they stop being resent. An instruction given at the top of a very long conversation may simply no longer be present by the time it matters.
Telling them apart
Context window vs Memory
Context window
A hard ceiling on one request. Everything inside it is available in full; everything outside is absent. Resets when the conversation does, and cannot be added to except by making the request itself smaller.
A product feature that stores selected facts about you and re-supplies them later. Persists across conversations, is usually editable, and is deliberately small because whatever it supplies has to fit inside the window like everything else.
If it forgot something from ten minutes ago, that is the window. If it recalled your name from last week, that is the feature.
Context window vs Token
Context window
The container: how much can be held at once, expressed as a number of tokens.
The unit the container is measured in: one chunk of text, roughly three-quarters of an English word.
They are the same measurement at different scales, like litres and a tank. A window quoted in the hundreds of thousands is counting tokens, so knock off roughly a quarter for a word count.
Questions
- How much text is a 200,000-token context window?
- Roughly 150,000 words of ordinary English, using the rule of thumb that a token averages about three-quarters of a word. That is around three hundred pages, or a couple of shorter books. Dense material with names, figures or code runs shorter, because it breaks into more tokens per page, so treat any conversion as a planning estimate.
- Why does the assistant forget what I said earlier?
- Because the earlier messages no longer fit. Every turn resends the conversation so far, and once the total would exceed the window the oldest material is dropped or compressed. Nothing announces this, so the first sign is usually the assistant repeating a question you already answered or ignoring an instruction you gave.
- Should I choose a tool with the biggest window available?
- Only if your work genuinely involves long single documents. Bigger windows cost more per request and do not improve reasoning. For most everyday tasks a moderate window with well-chosen extracts outperforms an enormous one filled with everything, because relevant material near the edges is weighted more reliably than the same material buried in bulk.
- Does attaching a file use up the context window?
- Yes, and usually more than people expect. The file's text is inserted into the request, so a long attachment can consume most of the budget before your question is read. Where a product offers it, searching a document and supplying only the relevant passages is far more economical than attaching the whole thing.
- How do I tell when a conversation has outgrown its window?
- Watch for the assistant re-asking something you answered, dropping a formatting rule you set at the start, or contradicting an earlier statement. Those are the usual first symptoms. None of them announces itself, so the practical habit is starting a new conversation when the subject changes rather than waiting for the signs.
- Is the context window the same as the model's training?
- No. Training is what the model learned in advance and cannot be changed by you. The window is what it can be shown right now, for this reply only. A model can have read enormous amounts during training and still be unable to see a document you have not attached.
Key takeaways
- The window is the total text a model can consider for one reply, measured in tokens.
- Instructions, attachments, conversation history and the reply itself all share the one budget.
- Overflow is usually silent: the oldest material is dropped or summarised and the answer continues.
- Capacity is not attention, and material buried mid-input is weighted less reliably than material at the edges.
- A well-chosen extract usually beats an enormous attachment, even when the attachment fits.
Last checked July 2026