Very cool paper on memory compression for agents.

If you run many agent sandboxes in parallel for RL or evals, memory becomes highly redundant. This work suggests that compressing against that redundancy cuts sandbox memory by up to 8.7x.

Memory is becoming the capacity limit for high-fanout agent workloads.

One task can spawn many concurrent sandboxes, and they all start from the same template and run related trajectories.

HKUST researchers measured 76 to 96% of pages with template-relative or cross-sandbox redundancy.

AgentZip compresses pages against the template and against sibling sandboxes, including pages that are similar without being identical.

It runs expensive compression while the agent is waiting on the LLM, and it prefetches pages at restore time to control slowdown.

Results:

Sandbox-owned memory drops by up to 8.7x, against 2.1x for the Linux configuration. Aggressive compression slows execution by 3.1x on its own, and the scheduling and prefetching bring that down to 1.40x.

Paper: https://arxiv.org/abs/2609.11294

Chat with Paper: https://academy.dair.ai/papers/memory-compression-for-high-fanout-agent-sandboxes-2609.11294