AWhat changed
An agent may repeatedly make the same mistake even when a previous attempt exposed it. ReasoningBank turns self-judged successful and failed trajectories into short strategy records with a title, description and content. Later tasks retrieve relevant records, and new experiences add further lessons. Memory-aware test-time scaling, or MaTTS, compares parallel attempts or uses sequential refinement to obtain richer experience. Ouyang and colleagues test the combined approach on web navigation and software issues. The contribution is the particular strategy-memory and scaling method, not the general idea that failures can teach useful lessons.
Scientific scope
An agent may repeatedly make the same mistake even when a previous attempt exposed it. ReasoningBank turns self-judged successful and failed trajectories into short strategy records with a title, description and content. Later tasks retrieve relevant records, and new experiences add further lessons. Memory-aware test-time scaling, or MaTTS, compares parallel attempts or uses sequential refinement to obtain richer experience. Ouyang and colleagues test the combined approach on web navigation and software issues. The contribution is the particular strategy-memory and scaling method, not the general idea that failures can teach useful lessons. Success and failure labels come from a model and can be noisy. Reduced interaction steps do not include all memory preparation or token costs. MaTTS spends extra computation, and best-of-several results differ from a single attempt. Some web task success rates remain low despite better step-level metrics. Failure on one attempt does not show that a route is impossible. This is not verified autonomous lifelong learning, an adversarially safe memory service or a separate discovery for each tested model.
Evidence
The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. Success and failure labels come from a model and can be noisy. Reduced interaction steps do not include all memory preparation or token costs. MaTTS spends extra computation, and best-of-several results differ from a single attempt. Some web task success rates remain low despite better step-level metrics. Failure on one attempt does not show that a route is impossible. This is not verified autonomous lifelong learning, an adversarially safe memory service or a separate discovery for each tested model.
Contributors
- The named authors in the research record · SCIENCE: Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.
Provisional WRS assessment, edition 2026-10-23.9. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction