Paper Tests Agent-Controlled Forgetting To Cut Tool-Use Context

A preprint posted to arXiv on October 6, 2026 by Jan-Peter Franke tests a simple idea for agents that repeatedly call tools: let the model decide which of its own past tool outputs it no longer needs in full, swap each one for a short note at the same spot in the conversation, and keep the full original somewhere it can be recovered if needed later. The paper calls this agent-controlled forgetting, and it is implemented through a Python harness that lets the agent archive results in batches and recover them explicitly, with no model fine-tuning required. User instructions and the assistant’s own messages are protected from the process, so only tool observations get pruned.

The study is small and exploratory rather than a benchmark. It runs one OpenTelemetry debugging case followed by an unrelated implementation task, plus a separate application-development pair and an earlier continuation case for comparison. In the debugging case, the method ended with far fewer provider-reported prompt tokens than the version that retained full history, and used half as many cumulative input tokens overall. The author estimates the agent-controlled-forgetting run cost roughly a dollar less than retaining everything, though it issued more individual requests and took longer to finish. Both the forgetting arm and the retained-history arm passed what the paper calls the primary behavioral oracle, a two-case check of correct task completion, but neither satisfied a stricter follow-up evaluation.

The result that matters most for anyone considering the technique is not the saving itself but its unevenness. The application-development pair showed no context or cost saving at all, and an earlier continuation case actually produced lower manually assessed output quality despite using less context. The paper frames workload dependence as the central open question for reversible context management, not a solved problem.

The numbers

• 231,951 provider-reported prompt tokens with forgetting against 912,492 with full retained history in the debugging case.
• 50% fewer cumulative input tokens used overall with the method applied.
• An estimated USD 1.28 to 1.44 API cost against roughly USD 4.38 for the retained-history run.
• 17% longer run time for the agent using forgetting in that same case.

For teams running agents against noisy, tool-heavy trajectories such as log analysis or debugging, this is early evidence that pruning old tool payloads can meaningfully lower per-session token bills, provided the archive and recovery step is reliable enough that nothing important is actually lost. For flatter, more linear tasks, the paper’s own application-development result suggests the technique may add complexity without changing the bill at all, which argues against applying it as a blanket default across every agent workflow.