Proactive Memory Agent for Long-Horizon Tasks
Md. Shakhoyat Rahman Shujon · hf · 2026-07-11
This work focuses on behavioral state decay in long-horizon tasks: as the trajectory grows longer, the state information critical for decision-making gets scattered across the context or even pushed out of the window, causing the agent to "forget" at crucial moments.
Method: Proactive Memory Agent
- Leaves the original action agent unmodified
- Introduces an additional memory agent
- The memory agent continuously maintains a structured memory bank
- It decides when to inject a contextually relevant reminder to the action agent and when to remain silent
Experimental Conclusions
- This module can be directly attached to frontier action agents and existing harnesses as a plug-and-play solution
- It improves pass@1 on both Terminal-Bench 2.0 and τ^2-Bench
- Benefits both weaker and stronger action agents
- Ablation studies show that selective intervention outperforms passive retrieval, constant injection, advisor-only approaches, and general retrieval
Training Attempts
The authors also trained an open-weight memory policy:
- Utilized SETA data
- Employed SFT + GRPO
- Achieved higher validation rewards with partial transfer to Terminal-Bench
Related event: Meta Proposes Active Memory to Fix Long-Horizon Agent Decay(4 posts)→
More from coding & agent
- GitHub Next's Chopin prototype explores real-time multiplayer agentic planning — max__drake · 2026-10-27
- TMLR paper: enforced thinking doesn't reliably make LLM agents disclose key info to users — SharonYixuanLi · 2026-10-27
- LLM-written Rust library bbtree verified: speed claims hold, memory savings reverse — lmoroney · 2026-10-27
- Resin lets an agent build standalone LLM inference engines; vLLM and SGLang optimize DeepSeek kernels differently — AccBalanced · 2026-10-27
- How I let my cloud Grok Bot code on my Mac mini via Tailscale, no public ports — talkaboutdesign · 2026-10-27
- Context layout is a cost decision: one memory edit can bust your entire agent prompt cache — dscape · 2026-10-27