ReasoningBank Explained: How AI Agents Are Finally Learning to Remember (rewire.it)

Google Research announced ReasoningBank, a lightweight agent-memory framework designed to solve “agent amnesia” — the tendency for LLM agents to repeat the same mistakes because each task is treated as a blank slate. Instead of storing raw logs or rigid workflow recipes, ReasoningBank distills each completed trajectory (success or failure) into compact, human-readable memory items (title, context, failed approach, root cause, corrective strategy, applicability). A four-stage closed loop — Retrieve (semantic top-k recall), Execute, Judge (LLM self-evaluation), Distill — continuously harvests transferable heuristics without retraining. The approach explicitly treats failures as first-class data (40% of initial patterns), reducing token waste (estimated 20–40%) and improving production reliability by preventing repeated mistakes like hitting API rate limits or misusing site search. ReasoningBank also introduces Memory-Aware Test-Time Scaling (MaTTS) to deepen experience: Parallel MaTTS runs k diverse rollouts and uses contrastive analysis to distill robust patterns across successes/failures, while Sequential MaTTS iteratively refines a single trajectory to capture nuanced improvements. Technically, the system relies on semantic retrieval, LLM-based self-critique, and confidence updates (Bayesian-like reinforcement of memory strength) to generalize heuristics across contexts. The result is self-evolving agents that gain strategic wisdom over time—no fine-tuning required—making autonomous, multi-day, and production-grade agent deployments more reliable and cost-effective.