CMU & Meta Open-Source ACM: Agents Learn to Compress Context Autonomously, Cutting Peak Tokens 20%
aigclink · x · 2026-08-06
CMU and Meta open-sourced the Agentic Context Management (ACM) framework, addressing the context bottleneck in long-horizon agentic tasks.
Unlike traditional heuristic-based truncation or summarization, ACM equips models with two core tools:
- managecontext: Empowers the agent to autonomously decide when to compress context via post-training, replacing rigid external triggers.
- querymemory: Offloads compressed content losslessly to an external memory system, allowing on-demand retrieval later.
Researchers noted that frontier models like GPT-5.5 rarely use these tools natively. To solve this, they designed a bidirectional teacher-student distillation pipeline where a 397B model teaches a 9B model when to compress versus when to keep searching. The trained Qwen3.5-9B achieved a 27% relative improvement on BrowseComp-Plus—rivaling open-source models 40x its size—while reducing peak token usage by 20%.
More from coding & agent
- Cloudflare Launches AI Search with Free Embeddings and Reranking — michellechen · 2026-08-06
- Cloudflare Unveils Kitesurf, a Browser Built for AI Agents — threepointone · 2026-08-06
- Cloudflare Introduces Kitesurf: An Agent-First Browser Running on Workers — threepointone · 2026-08-06
- Cloudflare Introduces Fully Stateless Next-Gen MCP Architecture — threepointone · 2026-08-06
- Why We Tolerate AI Code Slop but Reject AI Writing Slop — bendee983 · 2026-08-06
- Sentence Transformers v5.7.0 released: fixes gradient caching, boosts training & inference speed — tomaarsen · 2026-08-06