Launching today
ReasoningBank by Google
Open-source memory framework for self-evolving agents
4 followers
Open-source memory framework for self-evolving agents
4 followers
ReasoningBank is an open-source agent memory framework that distills reasoning patterns from both successful and failed runs, helping agents improve continuously after deployment. For AI researchers and agent builders.
AI agents fail the same way twice. No existing memory system asks why.
ReasoningBank is a Google Research framework that distills structured reasoning strategies from both successful and failed task trajectories, giving agents a memory that compounds over time.
What it is: An open-source agent memory system that continuously extracts reusable insights from what the agent did, what worked, and critically, what did not.
The problem: Existing approaches either log raw action sequences (too noisy to reuse) or only capture successful workflows (which discards the richest learning signal). Neither produces transferable reasoning. Agents stay stuck in the same strategic blind spots.
What makes it different: ReasoningBank mines failures for counterfactual insight. An agent that got trapped in an infinite scroll loop does not just log the error. It distills a preventative strategy: verify the page identifier before attempting to load more. That is the difference between storing a checklist and building a mental model.
Key features:
Structured memory items with title, description, and distilled reasoning content
Continuous retrieval, extraction, and consolidation loop during deployment
Memory-aware test-time scaling (MaTTS) via parallel and sequential trajectory comparison
Who it's for: ML researchers and AI agent engineers building LLM-based agents for web navigation, software engineering, or any persistent multi-step task environment.
Results: 8.3% higher task success on WebArena, 4.6% on SWE-Bench-Verified, and roughly 3 fewer execution steps per task versus memory-free baselines. MaTTS compounds the gains further.