The longer an AI model's working memory, the more likely it is to just give up.
A new paper from researchers studying long-horizon search tasks documents what they call "context rot" — a degradation in model behavior as the amount of text in an AI's context window grows. Testing four major open-source models across three benchmarks, they found a consistent pattern: models fed extensive context increasingly either abandon tasks outright or hedge with uncertain answers before finishing. The problem compounds as context length increases, suggesting it isn't a random quirk but a structural weakness.
The finding matters because the industry has spent years treating longer context windows as an unambiguous win. Context rot suggests there's a ceiling — not on how much text a model can technically process, but on how reliably it acts on that text. For anyone building AI agents that browse the web, analyze documents, or complete multi-step research tasks, that's a real operational limit, not a theoretical one.
The researchers tested two mitigation paths: context management (seven methods across three categories, evaluated on performance, cost, and rot impact) and rejection sampling with a rot-aware filter. Both helped; combining them helped more. None of that changes the underlying finding — that the "bigger context is better" assumption the whole agentic AI boom is built on is shakier than the product demos suggest.