AgentZip compressed repeated memory across 16 isolated AI-agent workspaces.
Physical memory use fell by an average of 88.55%.
The same workloads took 40.3% longer to complete.
Running many AI agents in parallel creates a less visible bottleneck than model computation: each agent needs an isolated workspace for files, commands and test programs. Researchers at the Hong Kong University of Science and Technology have built a system that compresses the repeated memory inside those workspaces while the agents wait for their language model.
Platforms place an agent inside a sandbox so its commands cannot affect other jobs. When one problem is explored along 16 candidate paths, the platform may create 16 closely related sandboxes. They start from the same template and often load similar files and libraries, but small changes prevent Linux's conventional same-page merging from combining much of their memory.
AgentZip searches for candidate pages while an agent executes a command. Once the command finishes and the language model is preparing the next instruction, the system uses that idle interval to compress them. For each page it chooses among repeated-byte compression, storing the difference from a shared template, and a Zstandard dictionary learned from sibling sandboxes.
Compressed memory must be restored before a command reads it. AgentZip predicts the order in which pages will be needed from earlier access patterns, activity in related sandboxes and the stage of the tool sequence. It then uses Linux userfaultfd to restore pages in advance.
The researchers replayed software-engineering traces from 10 Python repositories on a single Intel Xeon server. Across 16 parallel rollout workspaces, AgentZip reduced sandbox-owned physical memory by 88.55% and increased completion time by 40.3%. Linux zswap reduced memory by 48.66% with a 43.6% time increase. Combining kernel same-page merging with zswap saved 51.24% and increased time by 51.7%.
The experiment covers CPU memory on one server and replays recorded activity. It does not reduce the GPU memory occupied by the language model, and it does not yet show user-facing tail latency in a live multi-tenant service. Even so, it identifies a resource that will matter as agent systems branch into more candidate plans: the duplicated computers surrounding the model can become as important as the model itself.
Sources
Memory Compression for High-Fanout Agent Sandboxes
Linux kernel documentation: Kernel Samepage Merging
Linux kernel documentation: zswap