摘要
AtMem Learned Memory Formation 是一次 memory write-path decision-model 实验。它公开了 offline agreement 改善但 sequential replay 漏掉全部 required stores 的负结果,并保留 deterministic formation 作为默认。(来源:Negative Learned Memory Formation: Not Worth Deploying Just for Offline Gain in AtMem)
核心内容
基本信息
- 实体类别:Agent memory write-path 实验与 checkpoint lineage。
- 关键名称或版本:Laya V2/V3、soft cross-entropy、RLCD、deterministic formation。
- 时间范围:论文于 2026-10-19 首次发布。
重要事实
- tuned V2 的 sealed agreement 为 73.79%,但 100-item operational replay 只有 16 个 exact decisions,required stores 为 0/25。(来源:Negative Learned Memory Formation: Not Worth Deploying Just for Offline Gain in AtMem)
- 数据严重 reject-heavy,且 train/serve packing policy 不一致;aggregate accuracy 因此不能代表 durable memory quality。(来源:Negative Learned Memory Formation: Not Worth Deploying Just for Offline Gain in AtMem)
关系与影响
它为 Harness Engineering 与长程 Computer-Use Agent 提供一条保守边界:learned writer 先 shadow、再 review,不应仅凭 offline aggregate metric 接管持久 memory mutation。
分歧与变化
论文没有主张 learned formation 普遍无效;结论只覆盖单一 encoder family、单 seed 与小型 inspected fixture。(来源:Negative Learned Memory Formation: Not Worth Deploying Just for Offline Gain in AtMem)
相关页面
- 概念:Harness Engineering、Computer-Use Agent、Execution-Based Agent Evaluation
- 实体:无
- 主题:Agent Harness 与 Agentic RL 开源生态
来源
待核实问题
- 需以独立、多 seed、正确 seed-state 的真实长程 Agent/CUA trajectories 验证。