AgentZip: Memory Compression for High-Fanout Agent Sandboxes
Abstract (opencli arxiv paper)
arXiv 2609.11294 abstract (opencli arxiv paper 2609.11294 -f json): AgentZip redesigns compression for AI agent sandboxes by exploiting within-template and cross-sandbox redundancy, extending compression to any page with a cheaper representation, switching from compression-time selection to restore-time prefetch, and aligning costly compression with LLM wait phases. Sandbox-private memory drops up to 8.7× (Linux baseline 2.1×); the worst-case performance penalty of aggressive compression drops from 3.1× to 1.40×.
论文要点 (中文)
HKUST + OPPO + Purdue（Mengming Li、Ceyu Xu、Qijun Zhang、Jiangnan Yu、Xiangfeng Sun、Haohui Mai、Zhiyao Xie）2026-09-10 提交。一个 agent 任务会派生出大量并发沙箱会话，从同一模板启动、执行相关轨迹，模板内和跨沙箱之间存在显著内存冗余。传统压缩在三个维度都不对：「怎么压」无法利用非相同页面的相似性；「压什么」保守选页只看 page-fault 开销；「何时压」被动按内存压力触发，无 agent 执行阶段感知。AgentZip 重做三个维度：双重机制利用模板相对+跨沙箱冗余；压缩对象扩展到任何「有更省表示」的页面；用 restore-time 预取取代压缩时的页面选择；把高成本压缩对齐到 LLM 等待期，避免与前台工具执行抢资源。在 LLM 训练/推理负载上，沙箱自属内存最多降 8.7×（Linux 2.1×）；restore 预取+agent 执行感知调度把激进压缩的性能惩罚从最高 3.1× 压到 1.40×，几乎留住全部内存收益。做 agent 平台/沙箱工程的读者必读——把「跨会话模板冗余」当成一等内存资源看，比让 OS 看不见地处理多省 4× 量级内存。
Key claims (English)
AgentZip (HKUST + OPPO + Purdue; Mengming Li, Ceyu Xu, Qijun Zhang, Jiangnan Yu, Xiangfeng Sun, Haohui Mai, Zhiyao Xie; submitted 2026-09-10) is a memory-compression system designed for AI agent sandboxes. It attacks all three weak spots of conventional memory compression: how (non-identical page similarity is wasted), what (overly conservative selection driven only by page-fault cost), and when (passive, on-memory-pressure, no agent-phase awareness). It exploits within-template plus cross-sandbox redundancy, extends compression to any page with a cheaper representation, switches from compression-time selection to restore-time prefetch, and aligns costly compression with LLM wait phases. On LLM training/inference workloads it drops sandbox-private memory up to 8.7× (Linux baseline only 2.1×); restore-time prefetch + agent-execution-aware scheduling reduce aggressive compression's worst-case performance penalty from 3.1× to 1.40×, keeping nearly all the memory win. For agent-platform/sandbox engineering, the takeaway is that treating cross-session template redundancy as a first-class memory resource saves roughly 4× more memory than letting the OS handle it opaquely.
Obsidian 证据摘录
「AgentZip 是专门给 AI agent 沙箱设计的内存压缩系统：利用模板相对和跨沙箱冗余的双重机制；把压缩对象扩展到任何『有更省表示』的页面；用 restore-time 预取取代压缩时的页面选择；把高成本压缩对齐到 LLM 等待期，避免与前台工具执行抢资源。在 LLM 训练与推理负载上，沙箱自属内存最多降 8.7×，Linux 配置只有 2.1×；restore 预取和 agent 执行感知调度把激进压缩的性能惩罚从最高 3.1× 压到 1.40×，几乎留住全部内存收益。」——OpenClaw定时任务/论文流水线/2026-09-15-论文流水线.md L25-27
链接
- 论文：https://arxiv.org/abs/2609.11294
- 论文流水线：OpenClaw定时任务/论文流水线/2026-09-15-论文流水线.md