The context window is the amount of text an AI chat can hold in view while it answers you, and once a conversation grows past it, the oldest messages drop out of sight. That is why ChatGPT can nail your first ten questions and then act like it never read the instructions you typed at the top.
The Gist
- A context window is the reading span the model keeps in view. Older messages fall out of it as the chat grows.
- Forgetting mid-conversation is normal behaviour, not a bug and not a sign your account is broken.
- Memory and the context window are two different things, and mixing them up is the most common beginner confusion.
- Three habits fix most of it: restate the essentials, start fresh chats on purpose, and attach long documents instead of pasting them.
Have ChatGPT Recap This Article
ChatGPTContents
What a context window actually is, in plain words
Why your instructions get dropped halfway through
Memory and context window are not the same thing
How to stop losing your instructions this week
What a context window actually is, in plain words
Picture a whiteboard behind the AI. Everything you type and everything it answers gets written on that board, and the model reads the whole board before writing its next line. The board has a fixed size, so once it is full, the oldest lines get wiped to make room for the new ones.
That whiteboard is the context window. It is not storage and it is not a filing cabinet. It is simply what the model can see right now, and nothing outside it exists as far as the answer is concerned.
The size is counted in tokens rather than words. A token is a chunk of text, often a short word or a piece of a longer one, and it is the unit these systems actually count. We walked through that unit in detail in our guide to what a token in AI really means for beginners, and it is worth a read if the word keeps showing up in your settings.
Sizes have grown a lot. OpenAI documents that its GPT-5.4 model supports up to one million tokens of context, which is enough for a stack of long documents at once. Older and cheaper models work with far less, and the free tier of any chat app rarely gets the biggest window on offer.
Here is the part almost nobody tells beginners. That window is shared, not reserved for you. OpenAI’s own help pages explain that the same space holds the system instructions, the tools available, any saved memories, and the model’s internal reasoning before it replies. Your conversation gets whatever is left.
The practical effect of that sharing is easy to underestimate. Turn on a stack of tools, keep a long list of saved memories, and ask the model to reason carefully before answering, and you have spent a slice of the window before typing a single word. The board was never empty when you arrived.
One more thing to unlearn: the window has nothing to do with your subscription being full or your account hitting a quota. It is a property of the model itself, the same way a page has a fixed number of lines. Upgrading a plan can get you a bigger page, though it never removes the edge of it.
Why your instructions get dropped halfway through
You set the rules at message one. Answer in French, keep it under 200 words, never use bullet points. By message thirty, bullet points are back and the answers run long. Nothing broke. Message one simply scrolled off the whiteboard.
The model does not rank your messages by importance. It reads what fits, and what fits is the most recent part of the conversation. An instruction you consider permanent looks exactly like a passing remark from three days ago once both are just text on the board.
Long pasted content burns through the window fastest. A single ten-page report can eat more space than fifty short questions, which is why a chat that felt sharp all morning turns vague right after you drop a document into it.
That said, there is a second mechanism at work, and it caught out plenty of users this month. When a conversation approaches the limit, some systems compress what came before instead of cutting it cleanly, and details get lost in the summary. We covered that behaviour when AI systems were caught quietly dropping user instructions in long chats.
Compression is the sneakier failure. A hard cut is obvious because the model plainly has no idea what you are talking about. A summary sounds confident and complete while having silently thrown away the one constraint you cared about.
There is a simple tell you can watch for. When answers start getting generic, hedging where they used to be specific, the model has usually lost the detail it was anchoring on. Vagueness in a chat that used to be sharp is almost always a window problem rather than a bad day for the AI.
Zooming out, this also explains a frustration plenty of people report with very long projects. A chat that runs for hours accumulates side quests, corrections and abandoned drafts, and every one of those lines competes for the same space as the brief you actually care about. Length itself becomes the enemy of precision.
Keep learning on AI Noobies:
- Your AI Forgets Instructions in Long Chats
- ChatGPT Trip Planning Works Until It Invents a Hotel
- AI Books Are 20% of Amazon’s Self-Published Catalog
Memory and context window are not the same thing
This is the confusion that costs beginners the most time. Memory and the context window solve different problems, and knowing which one failed tells you what to do about it.
The context window is per conversation and disappears when you close it. Memory is the feature that carries short facts about you from one chat to the next, the kind of thing that survives when you open a brand new conversation tomorrow morning.
OpenAI’s documentation is precise on this point. Saved memories work much like custom instructions, except the model updates them on its own rather than waiting for you to manage them by hand. Both memories and custom instructions then get loaded into the context of your next reply, which means memory spends part of your window too.
So the two interact rather than compete. Memory is the reason the assistant knows your job title in a fresh chat. The context window is the reason it forgets, inside that same chat, the outline you agreed on twenty messages earlier.
If the assistant keeps bringing up something about you that is out of date, that is a memory problem and not a window problem. Our step-by-step guide on how to see and erase what ChatGPT remembers about you covers the cleanup.
A quick diagnostic settles it in seconds. Open a completely new chat and ask the assistant what it knows about you. Whatever it answers came from memory, because nothing else survives the jump between conversations. Everything it cannot recall was living in the window you just left behind.
Custom instructions sit in a third category worth knowing. You write them once in the settings, they apply to every conversation, and unlike memory the model never edits them on its own. For a preference you want honoured permanently, that is a far more reliable home than hoping a chat remembers.
How to stop losing your instructions this week
Start with the cheapest habit. Every ten or fifteen exchanges on a long task, restate the two or three constraints that matter in a single short message. It feels redundant and it is the single most effective thing you can do, because it rewrites your rules onto the visible part of the board.
Second habit, and it goes against instinct. Open a new conversation on purpose when you switch topics. Keeping one endless chat running is what pushes your early instructions out, so a fresh window with a clean brief usually beats a tired one carrying eight hours of unrelated history.
Third habit, use attachments for anything long. ChatGPT already nudges you toward this: paste more than 10,000 characters into the message box and it converts the block into an attachment automatically, precisely so a large paste does not swallow your whole window.
For rules you truly never want to lose, put them in custom instructions instead of the chat. Those load with every conversation, which makes them the right home for things like your preferred language, your tone, and the formats you never want to see.
Fourth habit, keep your brief short and front-loaded. A three-line instruction stated cleanly survives longer and gets followed more reliably than a rambling paragraph that buries the same rules inside a story. Shorter rules take up less room, which literally means they stay visible longer.
On the flip side, resist the urge to dump everything you might need at the start of a chat just in case. Loading the window with material the model will not use crowds out the material it will, and it makes the eventual drift arrive sooner rather than later.
If a task genuinely needs a large document reviewed section by section, split it. Work through one part per conversation and paste your conclusions forward into the next. It sounds laborious, and it is still faster than arguing with an assistant that quietly lost page four.
Try one test today to see the effect for yourself. Take a long chat that has started drifting, paste your original three rules back in as a fresh message, and ask the same question again. If the answer snaps back into shape, you have just watched a context window get topped up, and you will never misread that failure again.
Frequently Asked Questions
Does ChatGPT remember previous conversations?
Not through the context window, which resets every time you open a new chat. It can carry facts across conversations only through the memory feature, which stores short details about you and loads them into later chats. If memory is switched off, each new conversation genuinely starts blank.
How many words fit in a context window?
It depends entirely on the model and the plan, and providers do not publish a simple word count for every tier. What is documented is the token capacity of the top models, with OpenAI listing up to one million tokens for GPT-5.4. Free tiers typically run on smaller windows than paid ones.
Why does the AI forget in the middle of a chat but not at the start?
At the start, everything you wrote still fits in the window. As the conversation grows, the earliest messages fall outside it and stop being visible to the model. Nothing is deleted from your screen, so the chat looks intact while the model is effectively reading only the recent portion.
Is a bigger context window always better?
A bigger window holds more, though it does not guarantee the model uses everything in it equally well. Long conversations can still get compressed or summarised, and details buried in the middle of a huge document sometimes get less attention than material near the beginning or end.
Test yourself
Your assistant forgot a rule you set twenty messages ago. Context window or memory?
Show answer
Context window. The rule scrolled out of the visible span inside that single conversation. Restate it in a new message and the model can see it again.
Your assistant greets a brand new chat by naming your profession. Context window or memory?
Show answer
Memory. The context window resets with every new conversation, so anything that survives across chats came from the memory feature.
You paste a 40-page contract and the answers immediately get vaguer. What happened?
Show answer
The document consumed a large share of the context window, leaving less room for your conversation. Attaching the file rather than pasting the text is the better move.
Stay tuned on AI Noobies.
About the Author
