Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents

This paper treats memory as an active intervention rather than passive retrieval for long-horizon agents. A separate memory agent updates a structured memory bank from recent trajectories and decides whether to inject a memory-grounded reminder into an otherwise unmodified action agent. The plug-and-play module improves pass@1 by 8.3 percentage points on Terminal-Bench 2.0 and 6.8 points on τ²-Bench across weaker and stronger action agents. Ablations report that selective intervention outperforms passive memory exposure, always-on injection, advisor-only guidance, and general retrieval. The authors also train Qwen3.5-27B with SFT and GRPO on SETA.