What it is
Training-free multi-agent framework that learns reusable procedures by exploring new software environments, then reuses frozen memory at test time. Lets an agent improve in unfamiliar digital environments without finetuning by autonomously exploring, verifying outcomes, and storing reusable procedures and failure lessons in persistent memory.
Key results
- OSWorld 2.0 mean partial-credit rises 71.97% -> 78.98% on 82 tasks
- ALE Near-term mean partial-credit rises 83.75% -> 84.82% on 67 tasks
- Keeps model parameters fixed across exploration and test-time execution
Shows that benchmark gains in new software environments are possible through verified memory reuse rather than weight updates.