Meta and CMU’s ACM lets agents edit and offload context for long tasks
omarsar0 · x · 2026-07-28
Meta and Carnegie Mellon researchers propose ACM (Agentic Context Management) for long-horizon agents.
Instead of using a rigid token threshold that compresses and discards context, ACM gives agents purpose-built editing tools so they can decide:
- when to compress,
- what to offload into external memory,
- and when to query that memory later.
The paper also introduces a post-training pipeline built from high-quality context-management demonstrations. Reported results include:
- 27% relative gain on BrowseComp-Plus
- performance close to open-source models 40× larger
The key claim is that better context management reduces peak token pressure, enables longer exploration, and improves consistency across trials.
Related event: Architectures Emerge for Long-Running Agent Context(4 posts)→
More from coding & agent
- ThePrimeagen mocks AI coding: infinite tokens still produce an Electron app — seatedro · 2026-10-17
- Generative UI 101: Building Live Interfaces with Agents on Opus 5.5 and Rivals — EXM7777 · 2026-10-17
- MCP server review: gate tool writes on auth/traversal findings, not test score — SamuelFmedeiros · 2026-10-17
- DeepSeek shows rapid vision gains: agent self-checks screenshots to build 3D scene — cedric_chee · 2026-10-17
- He gave an AI agent phone and email access to collect a $500 unpaid invoice — curious_vii · 2026-10-17
- AI:AM transcript: agent lockouts, hidden valence, and formal proofs for Linux kernel bugs — The Cognitive Revolution · 2026-10-17