DEVOURED
Proactive Memory for Long-Horizon Agents
Researchers achieved higher task pass rates by introducing a proactive memory agent that selectively reminds action-oriented models of critical, buried trajectory data.
What: The paper introduces a dual-agent architecture where a 'memory agent' runs alongside an action model, periodically injecting reminders into the context window to prevent 'behavioral state decay.' Testing on Terminal-Bench 2.0 showed a +8.3 percentage point gain.
Why it matters: This demonstrates a shift away from passive retrieval methods like RAG (Retrieval-Augmented Generation) toward active, agentic intervention to manage long-term state.
Deep dive
- The memory agent creates a structured memory bank from trajectories.
- It decides autonomously whether to intervene with a reminder or remain silent.
- This 'plug-and-play' module works with existing frontier models.
- It outperforms always-on context injection and passive retrieval strategies.
- The authors trained Qwen3.5-27B on the SETA framework to advance open-weight memory policies.
Decoder
- Behavioral state decay: The phenomenon where an AI loses track of environment facts or past decisions as a task trajectory grows too large for the context window.
- Pass@1: A metric representing the probability that a model solves a coding task correctly on its first attempt.
Original article
Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents
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 $\tau^2$-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 $\tau^2$-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.
Explore all 70 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi