Researchers introduce AgentZip, a memory compression system for AI-agent sandboxes that exploits redundancy across sandbox instances and between sandboxes and their shared templates. The system times expensive compression operations to align with LLM waiting periods rather than triggering solely on memory pressure, and uses restore-time prefetching to cut overhead. In evaluations, AgentZip achieves up to 8.7x memory reduction while limiting slowdown to 1.40x, versus 3.1x slowdown for naive aggressive compression.
