Google AI Proposes ReasoningBank: A Strategy-Level I Agent Memory Framework that Makes LLM Agents Self-Evolve at Test Time

Google Research has introduced ReasoningBank, a novel memory framework for large language model (LLM) agents that enables self-evolution by converting interaction tracesboth successes and failuresinto high-level, reusable reasoning strategies. Unlike traditional memory systems that store raw logs or rigid workflows, ReasoningBank distills experiences into compact, human-readable strategy items comprising titles, descriptions, and actionable principles such as heuristics and constraints, facilitating transferability across tasks and domains. Coupled with memory-aware test-time scaling (MaTTS), this approach significantly enhances agent performance, achieving up to 34.2%