Google researchers introduce ReasoningBank, an agent memory framework that distills reasoning patterns from both successes and failures to enable continuous post-deployment learning.
A team from Google published 'ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory' at ICLR, introducing a memory framework that stores structured reasoning insights — not raw action logs — from both successful and failed agent trajectories. The system operates in a closed loop: retrieve relevant memories, act, self-assess via LLM-as-a-judge, extract insights, and consolidate. Benchmarks on web browsing and software engineering tasks show higher success rates and fewer steps compared to trajectory memory baselines like Synapse and Agent Workflow Memory. Code is available on GitHub.
ReasoningBank solves a real architectural gap: agents running in production repeat the same strategic mistakes because they have no mechanism to extract transferable lessons from failed runs. Unlike trajectory memory that logs raw actions, ReasoningBank stores structured reasoning patterns — description plus distilled rationale — that generalize across tasks. The LLM-as-a-judge self-assessment loop means the memory system improves without human labeling.
Clone the ReasoningBank GitHub repo and run it against your existing agent's most common failure cases — measure whether the retrieved memory reduces retry rate on the same task class within your test suite.
Go to the ReasoningBank GitHub repo (linked in the paper abstract) and clone it locally
Tags