Agent-Controlled Forgetting Cuts Tool-Using AI Agent Costs by ~50% in New arXiv Study

It is a context-management method where the agent itself selects earlier tool results, replaces each with a short note in place, and stores the full original in an archive that can be explicitly recovered later.

In the OpenTelemetry debugging case it ended with 231,951 prompt tokens versus 912,492 under retained history — about 50% fewer cumulative input tokens — at an estimated $1.28–1.44 versus roughly $4.38.

The paper is exploratory and notes strong workload dependence; one task pair showed no savings at all. Teams should pilot it on noisy tool-use trajectories and measure quality, cost, and latency before scaling.

The method made more requests and took 17% longer. It produced no context or cost savings in an application-development pair, and an earlier continuation showed lower manually assessed quality despite reduced context.