The article introduces a proactive memory agent designed to address "behavioral state decay" in long-horizon tasks, where decision-relevant information is often lost as trajectories expand. This separate module runs alongside an unmodified action agent, actively updating a structured memory bank and injecting reminders when necessary.
- The system operates as a plug-and-play intervention mechanism that decides whether to inject memory-grounded reminders or remain silent.
- It achieves gains of +8.3 pp on Terminal-Bench 2.0 and +6.8 pp on $τ^2$-Bench for pass@1 across various action agents.
- Ablations demonstrate that selective intervention outperforms passive bank exposure, always-on injection, advisor-only guidance, and general retrieval.
- The authors trained Qwen3.5-27B on SETA using SFT and GRPO to create an early open-weight memory policy.
This approach improves performance by ensuring critical context influences decisions rather than being buried in the context window.