An AI that reminds only when it matters — the "Proactive Memory Agent" for long-horizon agents

This work (Proactive Memory Agent) gives long-horizon AI agents a mechanism that reminds them of what matters, exactly when it matters.

1A failure specific to long-horizon tasks

The underlying failure is specific to long-horizon tasks. The state a decision depends on — task requirements, environment facts, prior attempts, diagnoses, open subgoals — scatters across an ever-growing trajectory. As the trajectory expands, that state gets buried inside the context window or pushed beyond it, and stops influencing decisions at the moment it is needed. The authors name this failure mode behavioral state decay.

2Memory as intervention, not retrieval

The core idea is to design memory as an active intervention mechanism rather than passive retrieval. The action agent itself is left unmodified; a separate memory agent runs alongside it, continuously updating a structured memory bank from the recent trajectory, and deciding at each point whether to inject a memory-grounded reminder or remain silent. The module is plug-and-play with frontier action agents and existing agent harnesses.

3Results on Terminal-Bench 2.0 and tau2-Bench

Across Terminal-Bench 2.0 and tau2-Bench, the module improves pass@1 for both weaker and stronger action agents, with gains of +8.3 pp on Terminal-Bench and +6.8 pp on tau2-Bench. Ablations show that selective intervention outperforms passive bank exposure, always-on injection, advisor-only guidance, and general retrieval.

As an early step toward open-weight memory policies, the authors train Qwen3.5-27B on SETA using SFT and GRPO, improving validation reward and achieving partial transfer to Terminal-Bench.