# AgentZip：压缩并行智能体沙箱共享内存

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-09-19 09:25
- AIHOT 分数：34
- AIHOT 链接：https://aihot.news/items/cmu7q2us50mkbrogrg52dcmc2
- 原文链接：https://x.com/rohanpaul_ai/status/2101120427435446358

## AI 摘要

AgentZip 发现并行智能体 88.55% 的内存开销来自重复状态，通过压缩同源沙箱间共享的内存来降低占用。当单个任务启动多个沙箱时，它们从同一模板出发、运行相似代码与库，标准 Linux 工具要么单独压缩页面、要么只合并完全相同的页，无法覆盖这些重叠。AgentZip 在智能体等待 LLM 期间完成压缩，并在下次工具调用前恢复可能需要的页面。

## 正文

If 1 agent task launches many sandboxes, manage their memory together

AgentZip shows that much of the memory cost by 88.55% of parallel agents is duplicated state, so the fix is to compress what sibling sandboxes share during LLM waiting time.

When 1 task spawns many sandboxes, they are not truly independent. They start from the same template and often run similar code, files, libraries, and commands. That leaves a lot of repeated memory.

Standard Linux tools either compress pages separately or only merge exact matches, so they miss much of this overlap.

AgentZip compresses memory against the shared template or sibling sandboxes, does the heavy work while the agent waits for the LLM, and restores likely-needed pages before the next tool call needs them.