Distilling insights with ReasoningBank
ReasoningBank distills world reasoning patterns into high-level, structured reminiscences. Every structured reminiscence merchandise incorporates the next:
- Title: A concise identifier summarizing the core technique.
- Description: A quick abstract of the reminiscence merchandise.
- Content material: The distilled reasoning steps, determination rationales, or operational insights extracted from previous experiences.
The reminiscence workflow operates in a steady, closed loop of retrieval, extraction, and consolidation. Earlier than taking motion, the agent attracts upon the ReasoningBank to collect related reminiscences into its context. It then interacts with the surroundings and makes use of an LLM-as-a-judge to self-assess the ensuing trajectory and extracts success insights or failure reflection. Notably, this self-judgement doesn’t must be completely correct, as we discover ReasoningBank to be fairly sturdy in opposition to judgment noise. Throughout extraction, the agent distills workflows and generalizable insights from the trajectory into new reminiscences. For simplicity, we straight append these to the ReasoningBank, leaving extra refined consolidation methods for future work.
Crucially, in contrast to current workflow reminiscence methods that solely give attention to profitable runs, ReasoningBank actively analyzes failed experiences to supply counterfactual alerts and pitfalls. By distilling these errors into preventative classes, ReasoningBank builds highly effective strategic guardrails. For instance, as an alternative of merely studying a procedural rule like “click on the ‘Load Extra’ button”, the agent may study from a previous failure to “at all times confirm the present web page identifier first to keep away from infinite scroll traps earlier than making an attempt to load extra outcomes”.