Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note2 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 present a proactive memory agent for long-horizon tasks that helps an action agent remember decision-relevant state as trajectories grow. The memory agent updates a structured memory bank from the recent trajectory and decides whether to inject reminders when needed, improving performance on Terminal-Bench 2.0 and $τ^2$-Bench.

What the Document Says

  • The paper discusses a memory agent for long-horizon tasks where decision-relevant state can be buried in the context window or pushed beyond it, resulting in 'behavioral state decay'.
  • This memory agent is designed to run alongside an unmodified action agent and updates a structured memory bank from the recent trajectory.
  • It decides whether to inject a memory-grounded reminder or remain silent based on certain conditions.
  • The module is compatible with frontier action agents and existing agent harnesses.
  • Experiments show improvements in 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 demonstrate that selective intervention outperforms passive bank exposure, always-on injection, advisor-only guidance, and general retrieval.
  • The researchers also train Qwen3.5-27B on SETA using SFT and GRPO, aiming to develop open-weight memory policies.

Key Concepts Mentioned

  • long-horizon tasks
  • memory agent
  • behavioral state decay
  • structured memory bank
  • reminder injection
  • action agent
  • frontier action agents
  • agent harnesses
  • pass@1
  • $τ^2$-Bench
  • Terminal-Bench 2.0
  • ablation study
  • Qwen3.5-27B
  • SETA
  • SFT
  • GRPO

Claims Made by the Document

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

Evidence Used

  • Improvements in pass@1 on Terminal-Bench 2.0 and $τ^2$-Bench.
  • Results of ablation study.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

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.