Proactive Memory Agent Tackles State Decay in Long Tasks
metaresearch · hf · 2026-07-11
To address the "behavioral state decay" problem—where decision-relevant information gets drowned out as task trajectories grow longer—this work proposes using memory as an active intervention mechanism rather than a passive retrieval tool.
- Architecture Design: A standalone memory agent runs in parallel with an unmodified action agent. It updates a structured memory bank from recent trajectories and decides whether to inject a memory reminder or remain silent. This module is fully plug-and-play.
- Experimental Results: On Terminal-Bench 2.0 and τ²-Bench, it significantly boosted the pass@1 scores for both weaker and stronger action agents (+8.3 pp and +6.8 pp, respectively).
- Ablation Analysis: Selective intervention outperformed strategies like passive exposure, continuous injection, advisor-only, and generic retrieval.
- Model Training: As an early exploration into open-source memory policies, the team trained Qwen3.5-27B on SETA using SFT and GRPO, achieving higher validation rewards with partial migration 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