「睡眠计算」: Agent 运行痕迹的离线记忆巩固模式
- ID: 0019229b
- 原文链接: https://x.com/yibie/status/2075457839481708960
- 作者: yibie
- 日期: 2026-07-27
- 分类: agents
- 标签: agent-memory, sleeping-compute, memory-consolidation, offline-learning, agent-trace
- 质量评分: 4/5
中文摘要
yibie 提出的「睡眠计算」是 Agent 记忆系统的好抽象:把 Agent 运行时留下的检索、失败、人类修改等痕迹称为「尾气」,主张只在热路径抓原始记录,离线用 Generator → Reflector → Curator 三阶段提炼成可复用上下文。模式的好处是让学习和延迟拆开:用户等待时不交税,系统空闲时再变聪明。在线任务负责行动和低成本记录,离线任务负责压缩、反思和筛选。
English Summary
The 'sleeping compute' pattern treats agent runtime traces (retrievals, failures, human corrections) as 'exhaust' that should only be cheaply recorded on the hot path, then processed offline through a Generator → Reflector → Curator pipeline into reusable context. Learning and latency are decoupled: users never pay the cost during active sessions.
推荐语
Agent 记忆新模式:热路径只记录,离线 GeneratorReflectorCurator 提炼经验
来源: AK-RSS-Digest / X书签消化 2026-07-27