Context rot means an AI getting less accurate as you give it more and more text to work with in one go. The longer the conversation or the bigger the pile of documents, the more likely it is to miss or mix up details.
Picture a desk. With three papers on it, you find what you need at once. Pile two hundred papers on it and you start missing things, even though they are all technically “there”. Context rot is the AI’s desk getting too cluttered.
In a sentence: “This chat has been going since lunch and it’s started ignoring my rules; I think it’s context rot, so I’ll start a fresh one with a short summary.”
What the word means
The plain point: more text in front of the AI is not automatically better. Past a point, quality slips.
The context window is everything the AI can see at once: your messages, its replies, and any files you added. See context window meaning. Anthropic’s documentation names the problem: “As token count grows, accuracy and recall degrade, a phenomenon known as context rot.”1 A token is a small chunk of text; see token meaning. Recall here means finding and using a detail that is somewhere in the text.
A 2025 technical report published by the company Chroma tested 18 models and found the same pattern, even on simple tasks.2 That is one company’s tests, not proof about every model.
Where the word comes from
The name started as a throwaway line on a forum. On 18 June 2025 a Hacker News user called Workaccount2 suggested it while describing how a model’s answers get worse as a long conversation fills up with distractions and dead ends.3 The programmer and blogger Simon Willison picked the comment up the same day and credited it with coining the term.4
The following month the company Chroma used it as the title of its technical report, and Anthropic’s documentation now uses the term too.21
A simple example
The plain point: you will usually notice it in long chats.
- Early on: you tell the AI “always write in British English and keep it under 200 words”. It does.
- Hours later: after dozens of messages and three pasted reports, it starts writing “color” and running long. Your rule is still in the chat; it just carries less weight among everything else.
If a long chat starts forgetting your instructions, a fresh chat with a short summary often works better than repeating yourself.
Why it matters to you
The plain point: what you put in front of the AI matters as much as how much room there is.
- Curate, do not dump. Anthropic says choosing what goes in matters as much as the space available.1 Paste the relevant pages, not the whole archive.
- Start a new chat for a new task. Old, unrelated history is clutter. Token economics covers this, and the cost side too.
- A huge window is not a promise. A model that can hold a lot of text will not necessarily use all of it well.
Often confused with
- Running out of room: when the window is full, the tool may refuse, cut off, or drop the oldest messages. Context rot happens before that, while there is still room.
- Hallucination: the AI inventing things. Context rot can lead to mistakes, but the cause is too much material, not missing knowledge. See hallucination meaning.
- Stale context: some people, including pages on this site, also use “context rot” loosely for saved background notes going out of date. See reusable context beats reusable prompts. The documented meaning above is about length.
Questions people ask
What causes context rot?
There is no single agreed cause. Chroma’s tests found that models do not pay equal attention to everything in front of them, and the report notes that distractors (text that looks relevant but is not) are known to make results worse.5 A long chat full of side-tracks and abandoned ideas is full of distractors.
Does a bigger context window fix context rot?
Not by itself. A bigger window lets the AI hold more text, but Chroma’s report argues that the popular “needle in a haystack” test (hide one sentence in a long document and ask the AI to find it) made long inputs look like a solved problem when they are not.5
How do I avoid context rot?
Keep what the AI sees short and relevant: paste the parts that matter, and start a new chat (with a brief summary) when a task changes or a conversation drags on. Anthropic says the same: curating what is in the context is “just as important as how much space is available.”1
Learn more on this site
- Managing what the AI sees: context engineering.
- Conversations that get worse: loops that improve and loops that degrade.
- The basics: the AI glossary and AI explained for complete beginners.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What is context rot?
- The AI forgetting everything when you close the app
- An AI getting less accurate as you give it more and more text at once
- Saved files going bad on your computer
Show the answer
Answer: An AI getting less accurate as you give it more and more text at once. Context rot means quality slips as the conversation or pile of documents gets longer, so the AI misses or mixes up details.
2. How is context rot different from running out of room?
- Context rot happens before the window is full, while there is still room
- Context rot only happens once the window is completely full
- They are exactly the same thing
Show the answer
Answer: Context rot happens before the window is full, while there is still room. Running out of room happens when the window is full; context rot starts earlier, while there is still space.
3. If a long chat starts forgetting your instructions, what does the page suggest?
- Repeat the instructions in capital letters
- Paste in even more documents
- Start a fresh chat with a short summary
Show the answer
Answer: Start a fresh chat with a short summary. The takeaway says a fresh chat with a short summary often works better than repeating yourself.
Related terms: AI search engine · Chain of thought · Context engineering · Context window · Deep research · Few-shot prompting
- 1Anthropic, Context windows, undated documentation, read at source 1 Oct 2026“more context isn’t automatically better. As token count grows, accuracy and recall degrade, a phenomenon known as context rot.” ↑a“more context isn’t automatically better. As token count grows, accuracy and recall degrade, a phenomenon known as context rot.” ↑b“This makes curating what’s in context just as important as how much space is available.” ↑c“This makes curating what’s in context just as important as how much space is available.” ↑d
- 2Chroma (Hong, Troynikov, Huber), Context Rot: How Increasing Input Tokens Impacts LLM Performance, 14 Jul 2025, read at source 1 Oct 2026“models do not use their context uniformly; instead, their performance grows increasingly unreliable as input length grows.” ↑a“models do not use their context uniformly; instead, their performance grows increasingly unreliable as input length grows.” ↑b
- 3Hacker News, comment by Workaccount2, 18 Jun 2025, read at source 12 Oct 2026“They poison their own context. Maybe you can call it context rot, where as context grows” ↑
- 4Simon Willison, quotation from Workaccount2 on Hacker News, 18 Jun 2025, read at source 12 Oct 2026“Workaccount2 on Hacker News, coining ‘context rot’” ↑
- 5Chroma (Hong, Troynikov, Huber), Context Rot: How Increasing Input Tokens Impacts LLM Performance, 14 Jul 2025, read at source 12 Oct 2026“Models often have to deal with distractors as well, which has been shown to degrade performance” ↑a“Models typically perform well on NIAH, which has led to the perception that long-context is largely solved.” ↑b
The comparison on this page (the cluttered desk) and the long-chat example are this site’s own, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK