智能体
x.com ·
@elvis(@omarsar0)
HKUST 提出 AgentZip：面向高扇出 Agent 沙箱的内存压缩，最高降 8.7 倍
该帖介绍一篇关于 Agent 内存压缩的论文。针对并行运行大量 agent 沙箱（用于 RL 或评测）时内存高度冗余的问题，HKUST 研究者测得 76%–96% 的页面存在模板相对或跨沙箱冗余。AgentZip 将页面同时针对模板和同级沙箱进行压缩，可处理相似但不完全相同的页面，并在 agent 等待 LLM 时执行压缩、在恢复时预取页面以控制减速。结果显示沙箱自有内存最多下降 8.7 倍（Linux 配置为 2.1 倍）；激进压缩本身带来 3.1 倍减速，调度与预取将其降至 1.40 倍。论文编号 arXiv 2609.11294。
AI推荐理由 转述 HKUST 的 AgentZip 论文（arXiv 2609.11294），给出具体量化结果：沙箱内存最多降 8.7x（Linux 配置为 2.1x），激进压缩单独导致 3.1x 减速，经调度与预取降至 1.40x，并称测得 76–96% 页面存在模板相对或跨沙箱冗余。信息密度高、有可查论文链接，对高并发 agent 沙箱/RL 评测的内存瓶颈有实际参考价值；但为二手转述，未提供同行评审或复现证据，8.7x 等数字应视为论文自报结果。
78
正文
Very cool paper on memory compression for agents.
If you run many agent sandboxes in parallel for RL or evals, memory becomes highly redundant. This work suggests that compressing against that redundancy cuts sandbox memory by up to 8.7x.
Memory is becoming the capacity limit for high-fanout agent workloads.
One task can spawn many concurrent sandboxes, and they all start from the same template and run related trajectories.
HKUST researchers measured 76 to 96% of pages with template-relative or cross-sandbox redundancy.
AgentZip compresses pages against the template and against sibling sandboxes, including pages that are similar without being identical.
It runs expensive compression while the agent is waiting on the LLM, and it prefetches pages at restore time to control slowdown.
Results:
Sandbox-owned memory drops by up to 8.7x, against 2.1x for the Linux configuration. Aggressive compression slows execution by 3.1x on its own, and the scheduling and prefetching bring that down to 1.40x.
Paper: [arxiv.org/abs/2609.11294](https://arxiv.org/abs/2609.11294)
Chat with Paper: [academy.dair.ai/papers/memory-…](https://academy.dair.ai/papers/memory-compression-for-high-fanout-agent-sandboxes-2609.11294)
💬2🔄0❤️6👀893📊3