Tokens
Context rot
Context rot is the degradation in recall and instruction-following that occurs as a model’s context window fills up.
Models attend more strongly to the beginning and end of a context than the middle. A decisive fact buried in the middle of a long document is the one most likely to be missed, and the effect worsens as total occupancy rises.
The practical implication is that curating what goes into context matters as much as how much space is available — a point Anthropic makes explicitly in its own documentation.
In practice
The measured effect is about position, not just volume. Liu et al. found accuracy dropping sharply when the answer sat in the middle of a long input on models that handled the same fact fine at either end. The practical response is not a bigger window — it is putting the decisive material immediately before the question.
Common questions
Where should I put the most important information in a long prompt?
At the end, immediately before the question. Attention concentrates at the start and the end of a context, so material buried in the middle of a long input is the most likely to be missed.
Is context rot a real measured effect?
The position effect is documented — accuracy drops sharply when the answer sits mid-input on models that handle the same fact at either edge. The exact curve varies by model and task, so treat published percentages with caution.