As agents become more capable they are shifting from one‑off tasks to providing sustained assistance across successive user goals. This shift demands that agents quickly discover how unfamiliar environments operate, reuse what they have learned across tasks, and keep that knowledge up‑to‑date as environments evolve. Existing benchmarks have begun to study experience reuse and adaptation, yet they offer little insight into which parts of accumulated experience agents revise, retain, or lose. To close this gap we introduce ReScopeBench, a benchmark covering 18 interactive environments that tracks the lifecycle of experience. It separates structural experience (how an environment is organized and accessed) from rule experience (reusable principles and constraints that guide decisions across tasks). We adopt an experience‑first construction approach, building task streams around shared experience and inserting silent environmental changes to probe acquisition, reuse, and revision. Experiments on 20 models evaluate task performance while monitoring changes in saved experience throughout each stream. Findings show that retained experience improves success rates, but agents struggle to keep it current when environments change; even when the required experience is correctly recorded, tasks may still fail. The work highlights selective maintenance and effective use of experience as complementary challenges for building robust long‑lived agents.
Review: ReScopeBench offers a fine‑grained framework for assessing long‑term agents, exposing the bottleneck of experience updating and pointing to the need for mechanisms that maintain relevance of knowledge in evolving settings.