In long-horizon tasks, decision-relevant states are often scattered across an expanding trajectory, requiring the action agent to surface and act on them.
As trajectories grow, task requirements, environmental facts, prior attempts, diagnoses, and open subgoals can get buried in the context window or pushed beyond it, failing to influence decisions when it matters. We refer to this failure mode as "behavioral state decay."
We view memory as an active intervention mechanism rather than passive retrieval. A separate memory agent runs alongside an unmodified action agent, updating a structured memory bank from the recent trajectory and deciding whether to inject a memory-grounded reminder or remain silent. This module is plug-and-play with frontier action agents and existing agent harnesses.
Across Terminal-Bench 2.0 and $\tau^2$-Bench, it improves pass@1 for both weaker and stronger action agents, with gains of +8.3 pp on Terminal-Bench and +6.8 pp on $\tau^2$-Bench. Ablations show that selective intervention outperforms passive bank exposure, always-on injection, advisor-only guidance, and general retrieval.
As an early step towards open-weight memory policies, we train Qwen3.5-27B on SETA using SFT and GRPO, improving validation reward and achieving partial transfer to Terminal-Bench.
Blogger's Review: The proposed proactive memory agent effectively addresses information loss in long-horizon tasks, enhancing the decision-making capabilities of agents through efficient memory management, showcasing promising practical applications.