ReasoningBank: A Memory Framework for AI Agents Learning from Successes and Failures
Google/DeepMind
Google researchers propose ReasoningBank, a novel agent memory framework that distills generalizable reasoning strategies from both successful and failed experiences, enabling continuous learning after deployment. Evaluated on web browsing and software engineering benchmarks, it improves agent effectiveness and efficiency. The framework uses memory-aware test-time scaling (MaTTS) to leverage exploration trajectories.
Research scientists from Google Cloud, Jun Yan and Chen-Yu Lee, introduced ReasoningBank, a new agent memory framework presented at ICLR 2025. It addresses the limitation of current agents that fail to learn from past experiences after deployment. Unlike existing methods that store exhaustive action logs or only successful workflows, ReasoningBank distills high-level, transferable reasoning patterns from both successes and failures. Each memory item is stored as structured records containing scenario, instruction, reasoning strategy, and action sketch. The system operates in a closed loop of retrieval, extraction, and consolidation, using an LLM-as-a-judge for self-assessment. ReasoningBank also introduces memory-aware test-time scaling (MaTTS) that uses exploration trajectories to generate high-quality memories through contrastive and refinement signals. Evaluated on WebArena and SWE-Bench-Verified using Gemini-2.5-Flash, ReasoningBank outperformed baselines including Vanilla ReAct, Synapse, and Agent Workflow Memory in task success rate and efficiency. The researchers observed strategic maturity over time, where simple procedural rules evolved into complex, preventative logic structures.
- Сокращения
- ICLR = International Conference on Learning Representations — Международная конференция по обучающим представлениям
- LLM = Large Language Model — Большая языковая модель
- MaTTS = Memory-Aware Test-Time Scaling — Масштабирование во время тестирования с учётом памяти
- AWM = Agent Workflow Memory — Память рабочих процессов агента
Source: Google Research —
original