Glossary · Updated 2026-10-19
What is context rot?
Context rot is the gradual drop in a model's output quality as its context window fills with stale, repetitive or irrelevant text. Old tool output, abandoned approaches and long logs compete with the current task for attention. Prevent it by clearing between tasks, keeping inputs relevant, and compressing bulky results before they enter the window.
What causes context rot?
Every turn of an agent session adds text: file reads, test runs, failed attempts, corrections. The model must find the relevant part among all of it. As the share of relevant text falls, the model is more likely to follow an outdated instruction, reuse a stale value or lose track of the goal. The term is widely used by practitioners, for example in this Towards Data Science piece on Claude Code.
What are the signs of context rot?
- The agent repeats a fix it already tried.
- It refers to files or errors from an earlier task.
- It ignores a recent instruction in favor of an older one.
- Answers get vaguer as the session gets longer.
How do you prevent context rot?
- Reset between tasks. In Claude Code that is
/clear. - Summarize deliberately. Steer
/compactwith focus instructions, or write a short state file and start fresh. - Keep inputs relevant. Read file ranges, filter command output, paginate tool results.
- Compress bulky results by question. A hosted query-aware model such as SuperCompress keeps only the original lines of a large output that answer the current request, so less irrelevant text accumulates.
Is context rot the same as hitting the context limit?
No. The limit is a hard ceiling. Context rot can start well before it, since quality depends on how much of the context is relevant, not just how full it is. See LLM context window too long.
How do I try compression against context rot?
Install the CLI and let setup wire your coding agents:
npm install -g supercompress-proxy && supercompress setupFrequently asked questions
What is context rot?
It is the decline in LLM output quality as the context window fills with stale or irrelevant material that competes with the current task.
Does a larger context window prevent context rot?
Not by itself. A bigger window can hold more irrelevant text. Relevance of the context matters more than size.
How do I fix context rot in Claude Code?
Use /clear between tasks, steer /compact when continuing, and keep bulky tool output out of the window by filtering or compressing it.
Can compression cause problems of its own?
Yes, if it drops a line the task needed. Extractive compression keeps original text, which makes omissions easier to spot, but test on your own work.