Agent-controlled forgetting reduces tool-use agent context costs by up to 50% in noisy trajectories

A new arXiv preprint proposes agent-controlled forgetting as a way to manage context in tool-using agents by having the model replace prior tool results with short notes at their original positions while keeping the full originals recoverably archived. The method is implemented in a Python harness that supports batch archival and explicit recovery without task-specific model training, and it protects user instructions and assistant messages from modification.

In an exploratory OpenTelemetry debugging scenario followed by an unrelated implementation task, the method ended with 231,951 provider-reported prompt tokens versus 912,492 under a retained-history baseline. It used 50% fewer cumulative input tokens and had an estimated API cost of USD 1.28–1.44 versus approximately USD 4.38. Both approaches passed the two-case primary behavioral oracle, though neither fully satisfied the follow-up evaluation.

The agent-controlled forgetting approach required 17% more requests and took 17% longer than the baseline. In a contrasting application-development pair, the method produced no context or cost savings, and an earlier continuation showed lower manually assessed quality despite reduced context, indicating that benefits depend strongly on workload characteristics.

The authors conclude that substantial resource savings are achievable in noisy tool-use trajectories, but emphasize that workload dependence is a central consideration for reversible context management.