Memory Compression for High-Fanout Agent Sandboxes

TL;DR - AgentZip is a memory-compression system tailored to high-fanout AI-agent sandboxes, exploiting shared templates and similarities across concurrent sessions. It cuts sandbox-owned memory by up to 8.7× while limiting compression-related slowdown to 1.40×.

• Compresses non-identical pages using template-relative and cross-sandbox redundancy.
• Compresses any page with a profitable representation, shifting overhead control to restore-time prefetching.
• Schedules expensive compression during LLM waiting periods to avoid disrupting foreground tool execution.
• Outperforms the evaluated Linux configuration’s 2.1× memory reduction while retaining nearly all aggressive compression savings.