5 min read
An AI chat does not remember everything you have told it. It remembers what fits in its context window, and once you understand that one idea, a long list of strange AI behavior suddenly makes sense: the assistant that forgets your pricing halfway through a proposal, the summary that skips the first half of a document, the chat that gets slower and vaguer the longer you use it.
This guide starts from zero. By the end you will know what a context window is, how to spot when you have hit its limits, and five simple habits that keep your AI work accurate.
The plain-English definition
A context window is the amount of text an AI model can look at in one go while it writes a reply. Think of it as the assistant’s desk. Everything on the desk (your instructions, the conversation so far, any files you uploaded, and the reply it is writing) can be seen at once. Anything that does not fit on the desk is out of view.
The size of the desk is measured in tokens, which are small chunks of text, often a short word or part of a longer one. In English, a token works out to roughly three quarters of a word on average, so 1,000 tokens is around 750 words. Anthropic’s documentation on context windows explains it the same way: the window is the model’s working memory for a single conversation, and it is different from everything the model learned during training.
Why a small business owner should care
Modern assistants have large windows, big enough for a long contract or a few hundred emails. So why does this matter? Three reasons.
- Long chats crowd the desk. If you keep one conversation open for weeks, your early instructions get buried under hundreds of later messages. Depending on the tool, older material may be summarized, trimmed or simply given less attention.
- Attention is not even. Research and everyday experience both suggest models pay most attention to the beginning and the end of what they are given, and can skim the middle. A key detail on page 23 of 40 is the one most likely to be missed.
- A fuller desk costs more. On plans and tools that bill or limit you by usage, every reply rereads the whole conversation. A bloated chat uses up your allowance faster.
Context window vs memory vs projects
These three get mixed up constantly, so it is worth separating them.
- Context window: what the model can see right now, in this conversation.
- Memory features: facts some assistants save about you between conversations, such as your business name. These get added back onto the desk when relevant, so they still take up space.
- Projects (and similar features like custom GPTs or Gems): a saved set of instructions and files that is placed on the desk at the start of every chat in that project. Claude’s help article on what Projects are describes this as a self-contained workspace with its own knowledge.
None of these give the AI a bigger desk. They decide what goes on the desk first. For a side-by-side of the project-style features, see our comparison of Custom GPTs, Claude Projects and Gemini Gems.
Four signs you have hit the limits
- The assistant ignores a rule you set at the start (“always quote in euros”) and slips back to its defaults.
- It contradicts a decision you made together earlier in the same chat.
- A summary of a long document covers the opening and closing sections well but is thin or wrong about the middle.
- Replies get slower, more generic or start inventing details. That last symptom overlaps with hallucination, which we explain in our beginner’s guide to AI hallucinations.
None of these mean the tool is broken. They mean the desk is cluttered.
Five habits that keep the desk clear
1. One task, one conversation
Write the newsletter in one chat and the invoice reminder in another. Short, focused conversations are faster, cheaper and more accurate. This is the single most effective habit on the list.
2. Write a handoff note before starting over
When a long chat starts drifting, ask: “Summarize the decisions we’ve made, the open questions and the current draft in under 200 words.” Copy that into a new chat and continue. You keep the progress and lose the clutter.
3. Put permanent facts in a project, not in chat
Your services, prices, tone of voice and standard policies belong in project instructions. That way they always land on the desk first, in the position the model pays most attention to.
4. Feed the relevant pages, not the whole file
If your question is about the cancellation clause, paste the cancellation clause. If you must upload a long document, tell the assistant where to look: “The answer is in section 7.”
5. Repeat the critical instruction at the end
For long requests, restate the one thing that must not be missed as your final line: “Remember: prices in euros, and no discounts above 10%.” Because models weight the end of the input heavily, this simple repetition prevents a lot of errors.
A note on privacy
Everything on the desk is, by definition, being sent to the AI provider for processing. Before you paste a client’s full file into a chat to “give it context”, decide whether the assistant really needs all of it. Our comparison of locked chats, Temporary Chat and Claude’s incognito mode explains which settings keep that material out of history and training.
Your first step today
Open your AI assistant and find your longest running conversation. Ask it for a handoff note using the prompt in habit 2, start a fresh chat with that note, and compare the quality of the next reply. Most people notice the difference immediately. Once that clicks, the next useful idea to learn is MCP, the connector that lets AI read your other apps directly instead of relying on what you paste.
Which of the four warning signs have you seen in your own AI chats?
