Agent-Controlled Forgetting for Tool-Using Agents: Reversible Context Curation in Practice

Researchers studied agent-controlled forgetting, where a model replaces tool results with short notes while keeping originals in a recoverable archive. In a test case, the method used 50% fewer cumulative input tokens and reduced estimated API costs from approximately USD 4.38 to USD 1.28-1.44, though it took 17% longer and made more requests.