Google researchers introduced ReasoningBank, a novel agent memory framework designed to help AI agents learn from both successful and failed experiences during deployment. Unlike existing approaches that either record detailed actions or focus only on successful workflows, ReasoningBank distills higher-level reasoning patterns from both triumphs and mistakes, addressing a critical limitation that prevents agents from improving over time.

The framework operates through a closed loop of retrieval, extraction, and consolidation, creating structured memories with titles, descriptions, and distilled reasoning steps. Coupled with memory-aware test-time scaling (MaTTS), ReasoningBank generates multiple trajectories and extracts insights through contrastive signals, creating a synergy where enhanced exploration feeds back into more refined memories.

Evaluated on WebArena and SWE-Bench-Verified benchmarks, ReasoningBank outperformed memory-free baselines by 8.3% and 4.6% respectively, while reducing execution steps by nearly three per task. Researchers observed agents evolving from simple procedural checklists to sophisticated compositional logic structures, demonstrating the framework's ability to foster genuine strategic maturity in AI systems.