Proactive memory agent lifts long-horizon task pass rates by 8.3 pp on Terminal-Bench
A new memory module runs alongside unmodified action agents, deciding when to inject context reminders and when to stay silent, beating passive retrieval across two benchmarks.
A proactive memory agent architecture addresses what researchers call "behavioral state decay" — the failure mode in which task-critical facts, prior attempts, and open subgoals get buried or pushed out of context as trajectories lengthen in long-horizon tasks. The system runs a separate memory agent in parallel with an existing action agent, maintaining a structured memory bank and deciding in real time whether to inject a reminder or remain silent.
The module is plug-and-play with frontier action agents and existing harnesses. Across Terminal-Bench 2.0 and τ²-Bench, it improved pass@1 for both weaker and stronger action agents: +8.3 percentage points on Terminal-Bench and +6.8 pp on τ²-Bench. The authors tested the architecture against several baselines: passive bank exposure (the action agent sees the memory bank but the memory agent never intervenes), always-on injection (the memory agent injects at every step), advisor-only guidance (the memory agent replaces the action agent entirely), and general retrieval (a standard RAG setup). Selective intervention outperformed all of them.
Training and generalization
The team fine-tuned Qwen3.5-27B on a dataset called SETA using supervised fine-tuning and GRPO (a policy-gradient method). The resulting checkpoint improved validation reward and achieved partial transfer to Terminal-Bench, though full generalization remains an open problem. The preprint does not specify a release date for weights or code, but the work is positioned as a reproducible baseline for future memory-augmented agent research.
The core insight is that memory should be an active intervention mechanism rather than a passive store. The memory agent updates its bank from the recent trajectory and then decides whether the action agent needs a nudge. When it does intervene, it surfaces exactly the subset of state that matters for the next decision — task requirements, environment facts, prior failures, or open subgoals — rather than dumping the entire history or staying silent when context is missing.
