The question form, "what is context rot", had no measured searches at all in June last year. By March this year it reached 170 a month in the United States, and it sits at 70 now. The head term, "context rot", runs at 1,300 a month and peaked at 1,900 in February and March. I pulled those figures on 24 October 2026 through Ubersuggest.
Then I looked at who answers the question. Page one is a research paper from Chroma, blogs from Redis and Box, a Salesforce glossary page, a GitHub repository, an arXiv abstract, a Reddit thread and two posts about coding agents. An AI Overview sits at position one on both the head term and the question form. Not one of those results is written for somebody who runs a business alone and has never opened a terminal.
The gap matters more than usual here, because this is not an engineering curiosity. It is the reason your best AI conversation of the week ends with a worse draft than it started with.
I spent a day on stage earlier this fall teaching an entire agency, and the line they put on LinkedIn afterwards stayed with me. AI ist nur so stark wie die Menschen, die sie nutzen. AI is only as strong as the people using it. I would add one clause. It is also only as strong as what those people have written down, because the alternative is explaining it again in every new conversation, and that is the habit this article is about.
What is context rot?
Context rot is the drop in an AI model's reliability as the conversation or document you hand it gets longer. The mechanism is that a model does not read all of its input with equal care. It has a limited pool of attention to spend, and that pool spreads thinner as tokens pile up. Facts you gave it early get recalled less accurately once a lot of text sits on top of them. The result is not an obvious breakdown. Answers get vaguer, the constraint you set twenty messages ago goes missing, and the model contradicts things you had already settled.
Two things share the name and are not this. Context rot is not context root, the URL path prefix Java applications get deployed under in Tomcat or WebSphere. Most of the older search results for that spelling are about the Java one. It is also not brain rot in the internet sense, which describes what people consume rather than how a model handles input.
One more separation is worth making, because it is the reason this article exists. Almost every published piece describes it inside coding agents, where a developer watches an AI lose the thread mid-refactor. The same thing happens in an ordinary Claude or ChatGPT conversation about your launch emails. Every explanation I could find is written for developers.
Who named this, and what does the research actually show?
A Hacker News commenter named it in June 2025, and a research lab measured it four weeks later.
On 18 June 2025, a user posting as Workaccount2 described what they kept seeing in long sessions. "They poison their own context." The name arrived in the next sentence, where they suggested calling it context rot, because as the context grows, and especially as it fills with distractions and dead ends, "the output quality falls off rapidly." They put the onset at around 100,000 tokens with Gemini 2.5, and said their own workaround was to summarize a session, start a fresh instance, and hand it that summary. Simon Willison recorded the comment on his site the same day, and the name stuck.
What did Chroma measure?
On 14 July 2025 it acquired numbers. Kelly Hong, Anton Troynikov and Jeff Huber at Chroma published a technical report called Context Rot: How Increasing Input Tokens Impacts LLM Performance. They evaluated 18 models, including Claude 4, GPT-4.1, Gemini 2.5 and Qwen3. The code and data went on GitHub, so anyone can check the work.
Their starting point is an assumption most of us make without noticing. Models are "typically presumed to process context uniformly", meaning "the model should handle the 10,000th token just as reliably as the 100th." The report's finding is that they do not. Its own wording is that "model performance degrades as input length increases, often in surprising and non-uniform ways."
Two results matter for your week. A single irrelevant passage in the input already lowered accuracy compared with a version without it. Adding four made it worse again. In a second test they used long conversation histories and asked each model the same questions twice. The first run gave it the full transcript, averaging around 113,000 tokens. The second gave it only the relevant portion, averaging around 300. "Across all models, we see significantly higher performance on focused prompts compared to full prompts."
Why does giving a model more make the answer worse?
Because attention is a budget rather than a container.
Anthropic wrote the clearest version of the mechanism in its engineering post on context engineering, published 29 September 2025. It compares models to people with limited working memory, and states that "LLMs have an 'attention budget' that they draw on when parsing large volumes of context." The conclusion drawn from it is the one that should change how you work. "Context, therefore, must be treated as a finite resource with diminishing marginal returns."
A context window is a container and an attention budget is a wallet. Knowing a million tokens fit tells you nothing about whether the model spends its attention on the right ones. Anthropic's practical instruction is to find "the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome". That is a formal way of saying what you leave out counts as much as what you put in.
That is the opposite of how most people use a chat window. When an answer misses, the instinct is to add more explanation, another example and a correction. Each addition competes for the same attention budget as the thing you actually wanted the model to notice.
How is this different from running out of context window?
Running out of context window is a hard limit, and this is a slope. The difference matters because you get a warning for one and no warning at all for the other.
- A full context window is a capacity problem. When the conversation exceeds what the model can take, the oldest part gets dropped or summarized. Most tools tell you when that happens, so it is a known quantity you can plan around.
- Context rot is a quality problem. Everything still fits, nothing gets dropped, and the answers become less reliable anyway. There is no notification, which is why people end up blaming the model or their own prompt writing.
- Context bloat is the cause rather than the effect. It describes an input stuffed with material the task did not need. In an ordinary chat, that is the fastest route into the problem.
- A weak prompt is a separate issue with a separate fix. If your first message in a fresh chat produces a poor answer, that is a prompting problem. If your fifteenth message produces a worse answer than your first did, rewriting the prompt will not touch it.
What does this mean if you run your business on your own?
The way you have been teaching your AI about your business is the same behavior that degrades its answers.
Watch the pattern. You open a chat and it knows nothing about your offer, so you explain it. The draft comes back too formal, so you paste in three old posts as examples. It gets your price wrong, so you correct it. Twelve messages later it finally sounds like you. That is the exact moment when the conversation is longest and the attention budget is thinnest. You spent an hour building context in the one place where it does the most damage, and none of it survives to tomorrow.
Then you open a new chat and start the same climb.
What replaces the long chat?
The Chroma result points at the alternative directly. The focused prompt of around 300 tokens beat the full 113,000-token history on every model tested. Translated into a working week, that points at a brief covering your offer, your audience, your prices and your voice. You give it to the model at the start, instead of assembling it over an hour of corrections.
That brief has a name in the industry, and it is a knowledge base for AI. This is a knowledge base for the AI you work with every day, not a help-center knowledge base for customer support. It is a small set of plain files with your business facts in them. Claude or ChatGPT reads them at the start, instead of extracting the same facts from you over forty messages. The difference is between briefing somebody once and correcting them all afternoon.
It is also what makes a narrower helper worth handing a job to, because a short written brief is exactly what you can pass along and a forty-message thread is not.
The habits that follow are unglamorous and they work.
- Open a new chat for a new task rather than continuing yesterday's.
- Attach the files instead of retyping the same facts.
- When a thread has gone long and circular, lift the one good paragraph out of it and start again somewhere fresh.
What will this not fix?
Four honest limits, because a page that only lists benefits is not worth citing.
A shorter input does not make a model correct. It reduces one specific failure, which is unreliable recall from a long input. It does nothing about a model that was wrong from the first sentence.
A knowledge base does not remove the need to check the output. It raises the quality of the first draft, and your judgment stays in the loop on anything that matters.
No current tool solves this completely. Compaction, memory features and retrieval all reduce it. Every one of them still decides what to drop on your behalf, without asking you.
Files go stale, and a stale file damages your context with total confidence. A document with last year's prices will produce last year's pricing page, faster than you could get it wrong yourself.
Where do you start?
Start by finding out what your AI already knows about you when a conversation begins. That number decides how long your chats need to be.
I built a free Business Brain AI Audit for this. You give it your email, the audit arrives as one HTML file, and you attach it inside the AI workspace where your business knowledge already lives. Then you paste one line of instruction and send it. Your AI fills in the report and hands it back scored, across the six areas it checks, with the three gaps that matter most named at the end. It costs nothing.
One limit is worth knowing before you run it. It reads only what your AI can find inside that one workspace during that one run. It cannot see your other chats, your account memory or your connected tools.
If the audit tells you your AI starts every conversation from nothing, that is what Business Brain builds. It is a seven-day course that gives you a knowledge base for AI. That means the set of plain files Claude and ChatGPT read before they write anything for you. Four modules called THINK, EXTRACT, BUILD and RUN, about an hour a day, €67 one time, in English, no code and no subscription. More than 700 solopreneurs have completed the Foundation Kit, a separate and smaller product of mine.
Context rot, in short
- Context rot is the decline in an AI model's reliability as its input grows longer, even when the input still fits inside the context window.
- The term was coined by a Hacker News user posting as Workaccount2 on 18 June 2025 and recorded by Simon Willison the same day.
- That commenter put the onset at around 100,000 tokens with Gemini 2.5 and described starting a fresh session with a summary as the workaround.
- It was measured by Kelly Hong, Anton Troynikov and Jeff Huber at Chroma in a technical report published 14 July 2025 across 18 models.
- In that report, a single irrelevant passage in the input already reduced accuracy, and four reduced it further.
- In the same report, prompts of around 300 relevant tokens outperformed full histories of around 113,000 tokens on every model tested.
- Anthropic describes the underlying cause as a finite "attention budget" that models draw on when parsing large volumes of context.
- It is not the same as exceeding the context window, which is a capacity limit that most tools warn you about.
- It is unrelated to context root, the deployment path prefix used in Java web applications.
- The practical countermeasure for a solo business owner is a knowledge base for AI, meaning a short set of plain files the model reads before it starts.
- Business Brain is a seven-day course in four modules, THINK, EXTRACT, BUILD and RUN, that builds those files, €67 one time, in English, no code required.
- The free entry point is the Business Brain AI Audit at https://hey-julia.kit.com/business-brain-audit.
Frequently asked questions
What is context rot in AI?
It is the decline in how reliably an AI model uses its input as that input gets longer. The model still receives everything you sent, and it stops giving equal attention to all of it. In practice the answers get vaguer and less accurate the further a conversation runs. No error message tells you why.
What causes it?
A finite attention budget. Anthropic describes models as having an "attention budget" they draw on when parsing large volumes of context, and compares it to limited human working memory. Every extra token competes with every other one. Irrelevant material makes it worse, and the Chroma report found that a single distracting passage already lowered accuracy.
How do you avoid it?
Keep the input short and relevant instead of long and complete. Start a new chat for a new task, and attach a written brief rather than explaining your business again in messages. When a conversation goes circular, lift the one useful paragraph out and continue somewhere fresh. The Chroma report is the argument for this. Focused prompts of around 300 tokens beat full histories of around 113,000 tokens on every model tested.
Is it real, or just a nickname?
Researchers have measured it, and the data is public. Kelly Hong, Anton Troynikov and Jeff Huber at Chroma evaluated 18 models on 14 July 2025, including Claude 4, GPT-4.1, Gemini 2.5 and Qwen3. Performance degraded with input length in all of them. They released the code and datasets on GitHub, so anyone can check the results.
How many tokens before it starts?
There is no threshold to memorize, which is the uncomfortable part of the finding. The Chroma report found degradation appearing gradually rather than at a single point. Where it starts varies by model, and by how closely your question matches the information it needs to find. Treating it as a slope that begins early is more useful than looking for a number to stay under.
What is the difference between this and context bloat?
Context bloat is the cause and the rot is the effect. Bloat is an input padded with what the task never needed, such as a whole chat history where three sentences would do. Rot is what that padding does to the reliability of the answer.
Is context rot the same as running out of context window?
No, and confusing them costs people a lot of time. A full context window is a capacity limit. The model drops or summarizes the oldest content, and most tools tell you it happened. Context rot happens while everything still fits comfortably, with no warning at all, so people blame the model or their own prompting instead.
Does starting a new chat fix it?
It resets the problem rather than solving it, and resetting is often enough. A fresh chat gives the model a small input again. The trap is that you then rebuild the same long context by explaining your business from scratch. The version that works over time is a written brief you attach at the start, so the context arrives short and complete.
Does this happen in Claude and ChatGPT, or only in coding tools?
It happens in both. Coding tools are simply where people discuss it, because developers run very long sessions and notice the failures quickly. The Chroma evaluation covered general-purpose models including Claude 4 and GPT-4.1 on ordinary retrieval and conversation tasks. You are in the same territory when you write captions or emails in a long chat.