Extract insights with ReasoningBank
ReasoningBank distills international reasoning patterns into high-level structured reminiscence. Every structured reminiscence merchandise comprises:
Title: A concise identifier that summarizes your core technique. Description: A quick abstract of the reminiscence merchandise. Content material: A reasoning step, resolution rationale, or operational perception extracted from previous expertise.
Reminiscence workflows function in a steady closed loop of acquisition, extraction, and integration. Earlier than performing an motion, the agent makes use of the ReasoningBank to gather related reminiscences into its context. Then, work together with the surroundings and use the LLM as a decide to self-assess the trajectory of outcomes and extract insights on successes and reflections on failures. Specifically, this self-judgment doesn’t should be completely correct, as ReasoningBank has been discovered to be very sturdy to judgment noise. Throughout extraction, the agent extracts workflow and generalizable insights from the trajectory into a brand new reminiscence. For simplicity, we’ll add these on to ReasoningBank and go away a extra subtle integration technique for future work.
Importantly, in contrast to current workflow memorization methods that focus solely on profitable executions, ReasoningBank actively analyzes failed experiences to supply counterfactual indicators and pitfalls. By distilling these errors into proactive classes, ReasoningBank builds sturdy strategic guardrails. For instance, the agent might merely say “[さらに読み込む]As an alternative of studying procedural guidelines like “Click on a button,” you may study from previous errors like “All the time test the present web page identifier first earlier than loading additional outcomes to keep away from infinite scroll traps.”
