ReasoningBank is a memory workflow designed to distill global reasoning patterns into high-level, structured memories. The system stores each memory item with a Title, Description and Content, where the content captures distilled reasoning steps, decision rationales or operational insights from past experiences.
The workflow runs in a continuous loop of retrieval, extraction and consolidation. Before taking action, the agent pulls relevant memories from ReasoningBank into its context, then interacts with the environment and uses an LLM-as-a-judge to self-assess the resulting trajectory. It then extracts success insights or failure reflection from that trajectory and appends new memories directly to ReasoningBank.
The approach differs from existing workflow memory strategies that focus only on successful runs. ReasoningBank also analyses failed experiences to surface counterfactual signals and pitfalls. The aim is to turn mistakes into preventative lessons and stronger strategic guardrails.
The source gives one example: instead of learning a procedural rule like “click the ‘Load More’ button”, the agent might learn from a past failure to “always verify the current page identifier first to avoid infinite scroll traps before attempting to load more results”.
The text says the self-judgement does not need to be perfectly accurate, and that ReasoningBank is robust against judgment noise. It also says more sophisticated consolidation strategies are left for future work.
Source: research.google.
Companies can share verified announcements through Newz9’s international press release submission page.
