ReasoningBank

ReasoningBank is an agent memory framework that distills reusable reasoning strategies from an AI agent's own past task attempts, including failed ones, rather than storing raw interaction logs or only successful runs. Siru Ouyang, Jun Yan, and colleagues at Google introduced it in "ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory" (arXiv:2509.25140, September 2025, accepted at ICLR 2026). After each task, the agent judges its own trajectory as a success or failure and extracts a compact "memory item" describing the general strategy, decision point, or mistake involved, instead of saving the full sequence of tool calls and outputs. At the next relevant task, the agent retrieves matching memory items to guide its actions, and the memory store grows with each new experience.

The framework's central claim is that failed trajectories carry information that success-only memory discards. They mark specific decision points where an agent's approach went wrong, and contrasting failed and successful attempts on similar tasks helps the agent isolate which choice mattered. The paper also introduces memory-aware test-time scaling (MaTTS), which runs multiple parallel attempts at a task at inference time and uses the resulting spread of successes and failures as a signal for building memory items, rather than treating each rollout independently.

On web-browsing benchmarks including WebArena and Mind2Web, and on the software engineering benchmark SWE-Bench-Verified, the authors report that ReasoningBank outperforms baselines that store raw trajectories or successful-only routines. It improves task success rates while also reducing the number of interaction steps an agent needs to complete a task. Google published an accompanying research blog post and released the code on GitHub.