A July 2026 paper proposes a Proactive Memory Agent architecture in which a separate memory module runs alongside an unmodified action agent to counter behavioral state decay during long-horizon tasks. Rather than passively exposing retrieved context, the memory agent maintains a structured knowledge bank and selectively injects reminders into the action agent’s context. Evaluated on Terminal-Bench 2.0 and tau-squared-Bench, the approach improves task performance by 8.3 and 6.8 percentage points respectively, and the authors also fine-tune an open-weight Qwen3.5-27B memory policy using supervised learning and GRPO.
