Meta's Proactive Memory Agent fixes context rot, lifting Claude Sonnet 4.5 from 37.6% to 45.9%
DeepLearningAI · x · 2026-09-25
Over long trajectories, action agents lose track of early mistakes as contexts get truncated or key info drowns in irrelevant tokens. Yifan Wu and colleagues at Meta AI propose a Proactive Memory Agent: a separate agent maintains structured notes (tools, past errors, key facts) and injects short reminders only at crucial moments, avoiding attention overload.
Built on Claude Sonnet 4.5, Claude Opus 4.6, and Qwen3.5-122B-A10B action agents, with memory agents using Claude Opus 4.6 or an SFT+RL-tuned Qwen3.5-27B on SETA command-line problems. Performance improved across all benchmarks, with Claude Sonnet 4.5 rising from 37.6% to 45.9%.
More from coding & agent
- Free Bots: A Persistent 3D City Where AI Agents Live, Work and Build — Daniel_Farinax · 2026-09-25
- Teaching a local Qwen 3.5 4B a new domain: CPT vs RAG, a four-phase experiment with evals — funJS · 2026-09-25
- MCP servers show letter placeholders: the new icons field is often ignored by clients — StructureRude9956 · 2026-09-25
- Mystery model Space Bunny turns hand-drawn sketches into playable 3D games — omarsar0 · 2026-09-25
- Stanford's CLM: an 8B System 1 model with 9x faster inference, new agentic coding SOTA — StanfordAILab · 2026-09-25
- Two-person AI pod delivered in one quarter what 5 engineers scoped for two — alex_verem · 2026-09-25