Memento：面向长上下文处理智能体的上下文管理学习
原标题：MEMENTO: Learning Context Management for Long Context Processing Agents
AI 摘要
研究提出 Memento，一个基于 Qwen3-8B 训练的长上下文处理智能体，通过上下文压缩（compaction）学习管理固定大小上下文窗口。方法上先用合成轨迹做 SFT 引导记笔记行为，再用结合答案正确性、笔记质量启发式和置信度校准的 RL 奖励强化。在书籍级问答四个基准（最长 250 万 token）上，Memento 将 Qwen3-8B 准确率从 36.9% 提升到 67.8%，上下文始终不超过 20K token，成本低于 GPT-5.5 智能体的 1/50，接近约 100 倍大的 GLM-5.3 智能体（69.3%）且成本仅为其 1/9。
正文节选
Memento: Learning Context Management for Long Context Processing Agents Abstract LLM agents that operate over long inputs must perform context management: deciding what to retain in a fixed-size context window. A common approach is compaction, where the model replaces its history with a self-generated note. Yet off-the-shelf agentic models, including GPT-5.5 and GLM-5.3, often write ineffective notes, and reinforcement learning (RL) alone rarely discovers useful note-taking behavior. We present