Meta 论文提出让 AI 智能体自行管理上下文，无需硬编码遗忘规则，模型能判断哪些内容仍重要并保留，效果更好且算力更省。在深度研究基准上，该方法比 Codex 式摘要得分高 11.4%，算力消耗少 21.5%，且无需训练、现有模型即可实现。
Meta dropped a great paper.
AI agents perform better and cost less when they manage their own context, so stop hard-coding the rules for what they keep.
Right now, most AI agents follow rules you write to decide what to forget when their context fills up, like "summarize everything when it's full."
This paper shows you can delete those rules and let the model clean up its own context, and it does a better job with less compute, even with today's models and no training
Since the model understands the task, it can judge what still matters, keep exactly that, and throw out the rest.
On a deep-research benchmark, it scored 11.4% higher than Codex-style summarization while using 21.5% less compute.
For long-running agents, let a strong model manage its own context before writing more cleanup rules.
‼️The Bitter Lesson for context management: Giving LMs unrestricted control over their context beats human-designed SOTA! Introducing 🩵Context Language Models (CLMs)🩵 - Natively manage their own context - Treat context as a file - Learn policies in CLM weights, no harness在 X 查看被引用的帖子
来源：Rohan Paul · x.com