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 module runs alongside an unmodified action agent, actively updating a structured memory bank and injecting reminders when necessary.
- The system improves pass@1 by +8.3 pp on Terminal-Bench 2.0 and +6.8 pp on $\tau^2$-Bench for both weaker and stronger action agents.
- Selective intervention outperforms passive bank exposure, always-on injection, advisor-only guidance, and general retrieval in ablation studies.
- The authors trained Qwen3.5-27B on SETA using SFT and GRPO to create an early open-weight memory policy.
This approach demonstrates that active memory intervention is more effective than passive retrieval for maintaining context in complex agent workflows.