GLM-52 897
GPT-56SC 873
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861
GLM-5 856
MUSE-SPK 841
QWEN-38X 824 -2.3%
GPT-6A 820
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531
CL-OP46 497
CL-OP48 490 -0.2%
GLM-52 897
GPT-56SC 873
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861
GLM-5 856
MUSE-SPK 841
QWEN-38X 824 -2.3%
GPT-6A 820
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531
CL-OP46 497
CL-OP48 490 -0.2%
← Back to feed

Meta and CMU Turn Context Compression Into an Agent Action: 57.0% to 72.7% on BrowseComp-Plus

Meta and CMU’s Agentic Context Management paper attacks one of the least glamorous failure modes in long-horizon agents: bad memory housekeeping.

Most production agents compress context when they cross a token threshold. That is operationally convenient and behaviourally crude. A threshold can fire in the middle of a search, after the agent has assembled a fragile reasoning state, or right before it needs to use a detail that is about to be abstracted away. Context size becomes the trigger, not task state.

ACM changes the trigger. Compression becomes an action the agent can choose.

What ACM Adds

The method gives the agent three memory-control behaviours:

BehaviourWhat It Does
CompressReplace old turns with a short summary when the agent decides they are no longer needed verbatim
ArchiveStore raw messages outside the active context instead of deleting them
RetrieveQuery archived memory when a dropped detail becomes relevant again

The point is not to make context infinite. The point is to stop treating context management as a background garbage collector. In ACM, the agent decides when memory should move between active working context and external storage.

That maps more closely to how long tasks actually fail. Agents do not only run out of tokens. They also compress the wrong thing, keep the wrong thing, repeat searches because prior evidence disappeared, or answer early because the useful thread fell out of view.

The Benchmark Result

The headline number is a 27% relative gain on BrowseComp-Plus using Qwen3.5-9B.

SetupBrowseComp-Plus Pass@1Peak Context
ReAct baseline57.0%63K tokens
ACM post-training72.7%54K tokens

That is the useful shape of result: better task success with less peak context. The model is not brute-forcing with a larger window. It is using the window more deliberately.

The paper also reports more consistent responses across four independent trials and longer exploration before failure. That second point matters. Long-horizon agents often collapse not because they cannot solve a subproblem, but because they prematurely settle after losing track of the search space.

The Post-Training Trick

ACM is trained with demonstrations of good context management timing. A teacher process inserts context-management actions when rollouts begin looping, and removes premature compression when the better move is another search, document fetch, or final answer.

That detail is more important than the memory API itself. Retrieval-augmented memory systems already exist. The hard part is teaching a model when memory action is warranted. “Compress at 80% of context” is easy to implement and often wrong. ACM tries to learn the timing policy.

The result sits in the same family as retained reasoning, context compaction, and tool-specific post-training. Frontier agent performance is increasingly determined by the harness surrounding the model: what state persists, what gets summarised, what the model can inspect, and which operations are exposed as actions.

Why This Matters for Production Agents

Long-context windows did not remove memory management. They made memory errors harder to notice. A 1M-token context can still contain the wrong 1M tokens, and a long agent trace can still bury the decision that matters under thousands of irrelevant tool calls.

ACM’s contribution is to move memory from infrastructure policy into model behaviour. That will not create capable agents by itself. A weak agent with better compression is still weak. But for models already capable of long search, coding, browsing, or research tasks, memory timing becomes a performance lever.

The broader direction is clear: the next generation of agents will not just call tools. They will manage their own operating state, including what they remember, what they forget, and when they should go back to the archive.