Home / What is context rot? Why long AI chats get worse

What is context rot? Why long AI chats get worse

If a long chat with ChatGPT, Claude or Gemini starts forgetting rules, repeating mistakes and drifting off topic, that is not your imagination. It has a name: context rot.

Context rot in one paragraph

Context rot is the decline in a model's reliability as the amount of text in its context window grows. Chroma's technical report Context Rot: How Increasing Input Tokens Impacts LLM Performance (July 2025) tested 18 models, including GPT-4.1, Claude 4 and Gemini 2.5, and found that models do not process long inputs uniformly: their reliability declines as input length grows, even on simple tasks.

Why a bigger context window does not solve it

A larger window lets you fit more text. It does not guarantee the model uses all of it well. A separate 2023 study, "Lost in the Middle" (Liu et al.), found models use information at the start and end of a long input better than information in the middle. In a chat, the early rules and decisions end up far from the newest message.

What it looks like in a real chat

• It forgets rules or requirements you gave early on.
• It contradicts something it said earlier.
• It brings back suggestions you already rejected.
• Corrections do not stick.
• Replies get slower or more generic.

How to avoid it

1. One topic per chat. Start a new chat when the subject changes.
2. Restate what matters. Repeat key constraints just before you need them.
3. Watch the fill level. Switch at about 50 to 70 percent. ContextMeter shows it live on ChatGPT and Claude, and Gemini support is coming soon.
4. Hand off, do not paste. Use the handoff prompt to carry a short brief into a fresh chat.

More: when to start a new chat, fixing "conversation too long", and the free token counter.

Frequently asked questions

What is context rot?

Context rot is the drop in an AI model's reliability as the amount of text it has to read grows. Chroma's 2025 research tested 18 models, including GPT-4.1, Claude 4 and Gemini 2.5, and found they do not handle long inputs uniformly.

Is a bigger context window the fix for context rot?

No. A bigger window lets you fit more text, but the research shows reliability can still fall as input length grows. Keeping chats focused and starting fresh with a summary works better.

How do I avoid context rot in ChatGPT and Claude?

Keep one topic per chat, put key instructions and facts early and restate them when they matter, and switch to a fresh chat with a short summary at about 50 to 70 percent of the window.

What is 'lost in the middle'?

A 2023 study (Liu et al., published in TACL) found that language models use information at the start and end of a long input better than information buried in the middle.

Last reviewed 2026-10-08. Numbers change often; see the sources linked on this page.