Context Rot: Why Claude Code Sessions Decay, and How to Govern Them

The performance of large language models can degrade over long sessions, a phenomenon known as "context rot." This decay is caused by both intrinsic rot, which stems from the model's architectural limitations like the attention mechanism, and content rot, which is the accumulation of incorrect or irrelevant information. Intrinsic rot leads to a gradual erosion of performance as the context window fills, because even irrelevant information consumes part of a fixed attention budget. Content rot occurs when failed attempts, debugging detours, or overly broad searches add confusing or contradictory data to the session history. While intrinsic rot is a fundamental constraint, users can actively manage content rot by curating the context to maintain the model's effectiveness.