AgentZip could cut AI agent sandbox memory by up to 8.7x

A post summarizing HKUST research describes AgentZip, which targets redundant memory in parallel agent sandboxes—the environments agents run in. It says researchers found 76–96% of memory pages had redundancy relative to a shared starting template or other sandboxes. AgentZip compresses similar as well as identical pages against those references. The post reports up to an 8.7x reduction in sandbox-owned memory, versus 2.1x for the Linux configuration. Aggressive compression alone slowed execution by 3.1x; running expensive compression while agents wait on the language model and prefetching pages during restore brought that slowdown down to 1.40x.