Tag
This paper investigates how LLM agents lose plan information as it gets evicted from context during long interactions. Using replay pairing and compression stress tests, the authors show that standard agents do not carry plans as persistent state, and propose diagnostics to measure plan signal decay.