ReScopeBench Tests Experience Retention Across 18 Environments and 20 Models

ReScopeBench constructs task streams around shared experience in 18 environments and inserts silent changes to isolate revision failures. It separates structural experience on environment organization from rule experience on cross-task constraints. Evaluation tracks saved experience state and task completion for each stream across 20 models.

Data indicate retained experience correlates with higher success on repeated tasks, yet agents leave recorded experience unchanged after environmental shifts. Even when the required experience remains correctly stored, task completion drops. Selective maintenance and application emerge as distinct failure modes.

The benchmark directly addresses gaps in prior agent evaluations that measure adaptation without tracing which experience persists or decays. In security and healthcare domains, where environments evolve continuously, unmaintained experience creates persistent operational risk. Future agent designs must incorporate explicit revision mechanisms rather than rely on implicit reuse.

Deployment pipelines will require experience-lifecycle metrics as standard acceptance criteria before scaling long-lived agents.