Context rot
Context rot is the gradual decline in a model's reliability on information as more tokens fill its context window.
Category
Industry concept
See all terms
What is context rot?
Context rot is the gradual decline in how reliably a model uses information as more tokens fill its context window. It isn't the same as running out of room: a request can sit comfortably within a model's token limit and still get a worse answer than a shorter version of the same request would, simply because more content is competing for the model's attention.
This matters because it breaks a natural assumption, that a model handles token ten thousand as reliably as token one hundred. In practice, performance grows less reliable as input length grows, even on simple tasks like finding one fact in a long passage, so length itself carries a cost.
How does context rot differ from a context window limit?
A context window limit is a hard ceiling: the model cannot accept more tokens than the window allows, and the request fails past that point. Context rot happens before that ceiling, inside a request that's within limits: as more tokens go in, the model's reliability on that content quietly degrades, so having room left in the window doesn't mean every added token is still being used well.
What causes context rot?
Context rot comes from how a model attends to its input: more content means more competing information to weigh, and relevant details can get diluted by everything else in the request, especially when the useful part sits in the middle of a long input rather than near the start or end. The effect appear on straightforward tasks, not just complex reasoning, which suggests it's a property of how models process long input generally.
Qontext's context agents work against this: instead of loading as much of the context repository as fits, they surface only the relevant information for a given task, keeping what actually reaches the model's context window small and targeted.