Meta & CMU Paper: Agentic Context Management Boosts Long-Horizon Task Performance by 27%
rohanpaul_ai · x · 2026-07-30
A new joint paper by Meta and CMU argues that long-horizon agents require control over working memory in addition to larger context windows. Traditional context compression based on token thresholds is suboptimal.
The paper proposes Agentic Context Management (ACM), with the following core ideas:
- Context management as an action: The agent decides when to compress context. Old turns are replaced with a short summary, while raw messages are stored outside the system rather than deleted.
- On-demand retrieval: If an archived detail is needed, it can be retrieved on demand via querymemory.
- Training pipeline: During post-training, a teacher inserts context management when rollouts loop and removes premature compression when another search or document fetch is the better move.
Experiments show this method decreases average peak context from 63K to 54K tokens. On the BrowseComp-Plus benchmark, Qwen3.5-9B increases its Pass@1 accuracy from 57.0% to 72.7%, a 27% relative gain.
More from coding & agent
- Connected steps are not a workflow: nobody designs the handoff — usehive · 2026-10-16
- Making Blender animations with Opus, Astra and Argon: how you prompt still matters — chongdashu · 2026-10-16
- Running PolyStrata + Cline coding agent on a $0-fancy laptop at ~90 tok/s — rpwoerk · 2026-10-16
- How do you attribute agent gains: model training or infrastructure? — jrxna · 2026-10-16
- Feed Claude your books, blogs and top engineers' code for claude.md — 'it now writes better code than me' — blaizedsouza · 2026-10-16
- Open-source /label plugin auto-names and colors Claude Code sessions in one step — hermanschutte · 2026-10-16