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
- Clarifying Agent Concepts: Differences Between MCP, Skill, and Tool Call — yangyi · 2026-07-30
- Workflow Share: Using Opus for Planning and Grok Subagents for Coding — kevinnbass · 2026-07-30
- Mnemos Engine: Building a Continuity and Accountability Framework for Autonomous Agents — RileyRalmuto · 2026-07-30
- 2100-Run Agent Benchmark: Grok Tops Value, GLM Nears Frontier — andreisavu · 2026-07-30
- Nadella Demos New Copilot Skills: Building Enterprise ROIC App with One Prompt — satyanadella · 2026-07-30
- Seeking Recommendations: Agent-to-Agent Gateway for Poison Message Prevention — jeffrschneider · 2026-07-30