Research on long context models found a U shaped recall curve: facts near the start and end of a long prompt get used while facts buried in the middle are ignored, even by models advertising huge windows. Distractors compound the problem, and degradation worsens when the target fact differs in wording from the question. Filling a context window to its limit therefore trades recall for reach.
How do you mitigate context rot?
Treat context as a budget rather than a dump. Retrieve fewer, higher signal chunks instead of stuffing documents wholesale, place the most important material at the start or end, and compact older turns as agent memory grows. This curation discipline is exactly what context engineering formalizes: shrinking focused context usually beats growing complete context.
Does context rot cause hallucinations?
It contributes to them. A model that cannot reliably find a fact inside its own context may fill the gap with a plausible invention, which surfaces downstream as hallucination. Long running agentic workflows are most exposed because their context accumulates across dozens of tool calls.