context rot

Context rot describes a decline in an AI model's ability to track relationships and surrounding context as the material supplied to it grows longer. The term, which concerns how effectively large language models use information within long inputs, gained wide currency after a 2025 report by the vector database company Chroma.

It differs from a context-window limit, which specifies how much input a model can accept: errors can occur even when the material fits within that limit.

This entry is based on AIPOST articles and widely known facts. If something is wrong, please send us a correction request.