[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-a79f20ac-131a-40a2-9914-eca1a2e7432e":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"a79f20ac-131a-40a2-9914-eca1a2e7432e","Proactive Memory Agent 让长程 LLM 不再「行为状态衰减」：Terminal-Bench pass@1 拉高 8.3 pp","长程 Agent 的关键决策信息往往散落在不断膨胀的轨迹中，被上下文窗口或超出窗口的部分活埋，等真要用时却拿不出来。Yifan Wu 团队在 arXiv:2607.08716 把这种失效模式命名为「行为状态衰减」(behavioral state decay)，并提出一个看起来朴素却反主流的解法：把「记忆」从被动检索改成立刻干预。\n\n核心思路是引入一个外挂的 Proactive Memory Agent，跟原 action agent 并行跑，从最近的轨迹片段里持续维护一张结构化记忆库，然后由它自己决定要不要、什么时候、注入哪一条「记忆提示」给 action agent——该沉默就沉默。论文强调 plug-and-play：原 action agent 不动、agent harness 不动，只多加一个会开口的旁观者。\n\n实验数据挺能说明问题：在 Terminal-Bench 2.0 和 τ²-Bench 上，pass@1 分别提升 +8.3 pp 和 +6.8 pp，无论 action agent 是较弱还是较强。消融实验更值得玩味：选择性干预 > 始终注入 > 被动检索 > advisor-only——「记得多」不如「记得巧」。训练侧用 Qwen3.5-27B 在 SETA 数据集上做 SFT+GRPO，验证奖励和 Terminal-Bench 迁移都给出正向信号。\n\n这条路线对 frontier agent 是个福音：与其把上下文窗口再扩一轮，不如在决策前多插一个会挑时机说话的「记忆管家」。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.08716","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm","2026-07-13T18:30:00Z","2026-07-13T18:19:52.326348Z","2026-07-13T18:19:52.326357Z",true,"agent",8]