What Is Context Rot?

Four failures and one fix

• Poisoning: the agent trusts a wrong note it wrote itself.
• Distraction: the agent's long history pulls it into loops, where it repeats a failed step.
• Confusion: tools and text it does not need lead it to the wrong choice.
• Clash: old and new versions of a fact sit side by side.
• The fix: take things out of the context, then check the result with a model that did not write it.

Context rot: a longer input, a less reliable answer

Chroma's context rot report, from July 2025, tested 18 models. The tasks were simple. One was to find a single sentence in a long document. Another was to copy out a list of repeated words. Even so, the report found that "model performance consistently degrades with increasing input length." The drops were often hard to predict.

That study measured retrieval, meaning whether a model can find a fact in a long input. We covered that side, and how far some newer models have pushed it back, in long context vs RAG in 2026. It is not solved everywhere. This guide is about agents, where most of the long input is the agent's own work.

An AI agent is a model that works in a loop. It reads the task, uses a tool, reads the result, writes a note, and goes again. Each turn adds text to its context, and nothing comes out unless someone builds a step to take it out. The four ways this goes wrong have names, from Drew Breunig's June 2025 essay "How Long Contexts Fail": poisoning, distraction, confusion and clash.

What fills an agent's context window

A context window is the most text a model can read at once. In an agent that runs for a long time, it fills with six kinds of text.

A person writes the first two on purpose. The other four pile up as the agent works, and each of the four failures below starts in that pile.

Context poisoning: the agent believes its own notes

Context poisoning is when a wrong fact gets into the context and the agent keeps using it as if it were true. Most often the agent wrote the wrong fact itself. It skims a file, writes down a wrong conclusion, and from then on reads that note as settled.

Google's Gemini 2.5 technical report describes a Gemini agent that plays Pokemon. The report came out in July 2025. Its name for what went wrong is "context poisoning." The agent's goals and summaries got "poisoned" with false facts about the game. That kind of error "can often take a very long time to undo," the report says. Until then, the agent chased goals it could never reach.

This is the problem from why AI can't check its own work, inside a single agent. The model that wrote the note is the one reading it back, so it has no reason to doubt it.

Context distraction: the history outweighs the training

Context distraction is when the agent's own history becomes a stronger guide than what the model learned in training. It stops planning the next step and copies what it did before. You see it as a loop, where the agent tries the same failed fix again and again.

The same Gemini 2.5 report saw this. Once its context grew very long, the agent "showed a tendency toward favoring repeating actions from its vast history rather than synthesizing novel plans," the report says. The report calls this anecdotal, since it is one agent in one game.

Context confusion: every extra tool is one more wrong choice

Context confusion is when text the agent does not need still shapes what it does. Tool lists are the clearest case. Every tool an agent can call adds a description to its context. Each tool the task does not need is one more way to pick wrong.

The number people quote here comes from Less is More, a paper accepted at the DATE 2025 conference. Varatheepan Paramanayakam and three colleagues posted it on November 23, 2024. They gave a small, compressed model, Llama 3.1 8B, one query from the GeoEngine test set. With all 46 tools in its context, it picked the wrong tool and failed. With 19 tools, it picked the right one and finished. Its 16K context window held all 46 tools, so space was not the problem. The authors put it down to "the large number of available options confusing the LLM."

From Table II of Less is More, posted November 23, 2024. The model is Llama 3.1 8B, compressed, on an Nvidia Jetson AGX Orin, a small computer that uses little power. Time is seconds to run the query. Power is watts drawn.

That result has a narrow scope: one query, on one small model, on a low-power device. The paper's main subject is saving time and power on hardware like that, so a large model on a server could behave differently. What carries over is the direction. A shorter list of the right tools got a right answer where the full list got a wrong one.

Context clash: old notes and new facts sit side by side

Context clash is when the context holds two versions of the truth and nothing says which one won. The agent wrote a plan, then learned the plan was wrong, and both are still in its context. Your correction sits next to the mistake it corrected.

A May 2025 paper from Microsoft and Salesforce researchers, LLMs Get Lost In Multi-Turn Conversation, measured something close to this. They gave models the same tasks in two ways. One way gave the whole task at once. The other gave it a piece at a time, over several turns of a chat. Over several turns, every top model they tested did worse, "with an average drop of 39% across six generation tasks."

The cause they found is the clash pattern. Models guess at the task too early and try to give a final answer too soon. Then they lean on that early guess. The guess stays in the context, and later facts have to fight it. In the authors' own words, "when LLMs take a wrong turn in a conversation, they get lost and do not recover."

The four failures side by side

Each failure starts with different text in the context. That tells you which fix to reach for first.

How to prevent context rot: take things out, then check from outside

Every failure above starts with something in the context that should not be there. So each fix takes something out.

Validation has a catch. If the model that wrote the note also checks it, you are back at poisoning, because the reader and the writer are the same. The check works when it comes from outside the agent: the source file, a test, a person, or a model from a different lab. We cover how to wire that step in which model runs your skeptic node.

A bigger context window does not fix context rot

A bigger window lets the pile grow longer before it runs out of room. It does nothing about what is in the pile. A wrong note is just as wrong in a huge window as in a small one.

In Chroma's July 2025 tests, performance dropped as input grew, across all the models it tested. Some newer models hold up much better at length, as our long context guide shows. Those gains are in finding text, and finding a note does not tell the model whether the note is true.

What context rot means if you write with ChatGPT or Claude

You do not need to run an agent to hit this. A long chat is a context too. A wrong fact the chatbot gave you early on stays in the chat, and it will reuse that fact later with the same confidence. An old draft you pasted in sits next to the new one.

• Start a new chat for each new piece, and paste in only the current draft and your sources.
• When the chatbot gets a fact wrong, note the fix and start a new chat, so the wrong version is gone.
• Before you publish, check the claims with a different model from the one that wrote them.

Frequently Asked Questions

Keep reading