Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

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

Source

arXiv

arxiv.org/abs/2607.08716v1

Read the source

Source

Plain-Language Brief

Researchers propose a proactive memory agent for long-horizon tasks, where decision-relevant information can get buried as trajectories grow. The memory agent updates a structured bank from the recent trajectory and decides when to inject reminders to improve decisions. This module is compatible with existing action agents and improves performance on Terminal-Bench 2.0 and $τ^2$-Bench.

What the Document Says

  • The paper introduces a proactive memory agent for long-horizon tasks.
  • The memory agent runs alongside an unmodified action agent and updates a structured memory bank from the recent trajectory.
  • The memory agent decides whether to inject a reminder or remain silent.
  • The module is compatible with frontier action agents and existing agent harnesses.
  • Improvements were observed on Terminal-Bench 2.0 and $τ^2$-Bench, with pass@1 gains of +8.3 pp and +6.8 pp respectively.
  • Ablations show that selective intervention outperforms passive bank exposure, always-on injection, advisor-only guidance, and general retrieval.
  • The researchers also trained Qwen3.5-27B on SETA using SFT and GRPO, improving validation reward and achieving partial transfer to Terminal-Bench.

Key Concepts Mentioned

  • proactive memory agent
  • long-horizon tasks
  • action agents
  • memory bank
  • reminders
  • ablation studies
  • Qwen3.5-27B

Claims Made by the Document

  • The proactive memory agent improves performance on long-horizon tasks.
  • Selective intervention outperforms passive bank exposure, always-on injection, advisor-only guidance, and general retrieval.

Evidence Used

  • Improvements were observed on Terminal-Bench 2.0 and $τ^2$-Bench with pass@1 gains of +8.3 pp and +6.8 pp respectively.
  • Ablations show that selective intervention outperforms other methods.

Methods / Approach

  • proactive memory agent
  • structured memory bank
  • selective intervention

Limitations Stated by the Authors

Not stated in source

Limitations Noted by the Agent

Agent interpretation — not from the source.

The model was not evaluated on a wide range of tasks or environments beyond Terminal-Bench 2.0 and $τ^2$-Bench.

Open Questions

  • How does the proactive memory agent perform on tasks outside of Terminal-Bench 2.0 and $τ^2$-Bench?
  • What are the potential drawbacks or unintended consequences of using a proactive memory agent?

Notes for Future Q&A

May include interpretation. Labeled where it does.

  • The paper focuses on improving performance in long-horizon tasks by introducing a proactive memory agent. The memory agent updates a structured bank from the recent trajectory and decides when to inject reminders to improve decisions.
  • The module is compatible with existing action agents and improves performance on Terminal-Bench 2.0 and $τ^2$-Bench, outperforming other methods in ablation studies.

Raw Abstract / Excerpt

In long-horizon tasks, decision-relevant state is often scattered across an expanding trajectory, while the action agent must surface it and act. As trajectories grow, task requirements, environment facts, prior attempts, diagnoses, and open subgoals can be buried in the context window or pushed beyond it, failing to influence decisions when needed. We call this failure mode 'behavioral state decay'. We study 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. The module is plug-and-play with frontier action agents and existing agent harnesses. Across Terminal-Bench 2.0 and $τ^2$-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 $τ^2$-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, we train Qwen3.5-27B on SETA using SFT and GRPO, improving validation reward and achieving partial transfer to Terminal-Bench.