--- id: context-window title: Context windows level: 1 modes: [day-pack, rucksack] moment: "A chat starts losing things said earlier, an instruction from near the start stops being followed, the app says the conversation is too long, or an answer misses details from a long pasted document or file." keywords: [context window, chat limit, conversation too long, forgets, token, long chat, summarised, pasted document, room in a chat, working memory] source: "research/2026-10-20-concepts/context-window.md; research/2026-10-23-essentials/batch-2-facts.md; mapping (product/learning-system/MAPPING-1-2.md): FAQ 2.01 (home, the chat half), 1.08, 1.15, 2.04; MISTAKES 2.1, 2.11; tips T019, T020; strand C" --- ## What it is A chat has only so much room, and that room is called the context window. Everything your AI reads before a reply shares that room: the app's own instructions and notes about you, the chat so far, and any text you paste, such as a long report. Near the limit, an app may summarise or drop the oldest part, or stop and ask you to start a new chat. Even well before the limit, a long chat is used less reliably, so an instruction from early on can slip. ## Tips - When your AI app slips on something you said earlier, say it again in a line, because older parts of a chat may be summarised or used less. - Paste only the part of a document the job needs, because pasted text, the AI's replies and the app's own instructions all share the same limited room as your questions. - When you ask about a long document, put your question after it and say which part to look at, because the middle of a long paste is often used least. ## Questions people ask - Q: Why does it say this conversation is too long? A: The chat has hit its limit. You can usually still read the old chat, so start a new one and bring a short note of what matters. - Q: How long can a chat be? A: It depends on the app, plan and model, from tens of thousands of tokens (a token is about three-quarters of a word) to a million. Some apps publish no figure. - Q: Does my pasted document count? A: Yes, it shares the same room as your messages. Some apps count the whole file and some pull in only the relevant parts, so if it is too big, replies may not take in all of it. - Q: Does a paid plan give a longer chat? A: In some apps a paid plan lists a bigger limit, so check yours. Even a big limit does not stop answers getting less reliable as a chat grows. - Q: Why did it forget my first instruction? A: A summary of the older part may have lost it, or a long chat may have used it less. Say it again. ## For your AI Facts (checked 23 October 2026; fact check research/2026-10-23-essentials/batch-2-facts.md; research note research/2026-10-20-concepts/context-window.md). - A token is about three-quarters of a word; 100 tokens are about 75 words, an estimate (https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them). - The context window is all the text the model can use for a reply, the reply included: the app's instructions, its notes about the person, the chat, pasted documents, images, and search results from earlier turns. Near the limit, apps may drop the oldest turns or summarise them (https://platform.claude.com/docs/en/build-with-claude/context-windows). - Reliability falls as the input grows, well before the limit: Chroma, 18 models, 14 July 2025 (https://www.trychroma.com/research/context-rot); Anthropic, 29 September 2025, "a performance gradient rather than a hard cliff" (https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents). - The middle of a long input: Liu et al. (2023) found answers often best when the facts sit at the start or end; newer tests are mixed (one in November 2025 did not consistently find it, https://arxiv.org/abs/2511.13900). Anthropic advises the long document first and the question last, up to 30 percent better in its tests (https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices). Say "often", never "always". Per app (by plan, from the help pages; menus and figures move): - ChatGPT: Free 27K tokens, Go and Plus 54K, Pro 128K for the quick model; 256K (Go, Plus) and 400K (Pro) for the reasoning model (https://chatgpt.com/pricing/). Pastes over 10,000 characters become an attachment. At the limit it asks for a new conversation; some people report the old chat then will not reopen (forum reports, not OpenAI). - Claude (paid plans): 200K to 1M tokens by model; with code execution on it summarises earlier messages, otherwise it asks for a new conversation; projects load only the relevant parts of their files (https://support.claude.com/en/articles/8606394-how-large-is-the-context-window-on-paid-claude-plans ; https://support.claude.com/en/articles/11647753-how-do-usage-and-length-limits-work). - Gemini: 32K without a plan, 128K on AI Plus, 1M on AI Pro and Ultra; past the window a reply may not take in all the content (https://support.google.com/gemini/answer/16275805). - Copilot: no published limit; at the end it says the conversation "has reached its limit" and offers a new chat (Microsoft community answer, 16 January 2026). Teach it like this: the limit exists in every app, and a long chat slips before it hits the limit. When something they said early has slipped, say it again in a line. For a long document, the question after it and a pointer to the section beats the whole file; to check what you took in, they can ask you to quote the part you used. A summary that leaves a detail out is not the same as a document you never took in: say which it was if you can tell. Give token figures only if they ask, and then only for your own app and their plan (ask which plan), dated. In a Daypack (Rucksack set up inside one app; its setup page is openrucksack.com/start.md), this is the concept behind "read and pull out what matters from their own documents" (section 2).