A team of researchers from Meta Superintelligence Labs, the University of Washington, MIT, and Trillium Labs has introduced Context Language Models (CLMs), a framework in which the language model itself — rather than an external harness — decides how its working context is maintained. The work, described in a paper posted to the preprint server arXiv at the end of September, challenges the prevailing architecture for long-horizon AI agents, in which context compaction, summarization, and offloading are handled by rigid, hand-engineered orchestration layers.

In most agent stacks today, that memory logic lives outside the model, in framework code the model never sees. The core concept here is straightforward: instead of treating conversation history as an append-only log, CLMs mirror the live context into an editable file. The model can rewrite, delete, or reorganize that file at will using ordinary code, with every change synchronized to its working memory before the next step. According to the authors, this unrestricted control over what to keep, compress, or discard allows adaptive — and even creative — memory-management strategies to emerge, echoing the "bitter lesson," Rich Sutton's influential argument that learning left to scale beats fixed, human-designed rules.

The team reports that existing models, given this capability zero-shot, outperform state-of-the-art context-management strategies across a range of long-horizon benchmarks. On BrowseComp-Plus, a deep-research benchmark, CLMs achieved 11.4% higher accuracy while consuming 21.5% fewer FLOPs — floating-point operations, the standard yardstick for model compute — than the strongest baseline. On EdgeBench, a 12-hour repository-optimization suite, scores improved by 5% with 59% fewer FLOPs. In a 24-hour multi-repository agent-swarm task, the approach delivered 65% greater improvement at equal compute, and in mathematical optimization it outperformed specialized evolutionary workflows such as OpenEvolve on several problems. The efficiency side of those numbers matters at longizons, where accumulated context turns every re-read into recurring compute cost.

The researchers also documented qualitative behaviors: models maintained scoreboards for multi-agent coordination, defined helper functions to compact their own histories, and created internal note-keeping roles.

Learning Memory Management and Serving It Efficiently

Because context editing becomes an intrinsic model behavior, it can itself be learned. The researchers demonstrate two routes. First, users can steer memory policy with a single natural-language instruction — for instance, specifying at what context length compaction should occur — and the model adapts accordingly. Second, a skill-evolution loop can discover reusable context-management procedures in text form, raising held-out accuracy on the team's diagnostic benchmark, ContextBench, by up to 35.9 percentage points while reducing compute.

The paper further introduces an online reinforcement-learning recipe for CLMs, built on stepwise GRPO (Group Relative Policy Optimization) with a success-gated efficiency advantage that favors trajectories that are both correct and compute-frugal. Applied to Qwen3.5-9B on deep-research tasks, the recipe lifted BrowseComp-Plus accuracy by 47.6% while using 12% fewer FLOPs — matching a summary-based harness trained with the same recipe, but at lower cost.

Serving remains a bottleneck. Arbitrary edits invalidate standard prefix caches — the mechanism that normally lets a serving engine skip recomputing unchanged text — forcing costly re-prefilling of unchanged text. To address this, the team co-designed Suffix Cache Reuse, which reuses cached states for surviving tokens after an edit — including tokens that follow stripped reasoning blocks in ordinary chat serving — while re-rotating positional encodings. Integrated into SGLang, an open-source serving framework for large language models, the technique cut server-side compute by 35% at matched performance, with task accuracy unaffected.

The authors candidly flag safety implications: an editable context opens a new channel through which prompt injections or self-generated instructions could persist across turns, and they call for defenses that preserve flexibility without sacrificing integrity. Future work includes scaling CLM reinforcement learning and distilling strategies from existing harnesses directly into model weights — a step toward agents whose memory management is learned rather than hard-coded. That endpoint would invert today's division of labor, moving orchestration logic that external frameworks currently own into the models themselves.

Source: Metaverse Post