ReasoningBank Teaches AI Agents to Learn From Failure
Google Research's new memory framework helps deployed agents extract generalizable strategies from both successful and failed task attempts.
Can AI agents actually get better at their jobs after they're deployed?
Yes, according to Google Research. A framework called ReasoningBank, published at ICLR and developed by researchers at Google Cloud, gives agents a structured memory system that distills lessons from both successful and failed task attempts into reusable reasoning strategies — rather than just replaying raw action logs.
The problem with existing agent memory
Current agent memory approaches fall into two broad camps. Systems like Synapse store exhaustive action-by-action trajectories, while others like Agent Workflow Memory capture summarized workflows drawn only from successful runs. Google's researchers argue both approaches share a fundamental flaw: neither extracts higher-level, transferable reasoning patterns, and fixating on successes alone discards what can be the most instructive data — the agent's own failures.
Without a mechanism to analyze past experience, an agent deployed on long-running tasks will repeat the same strategic mistakes indefinitely, treating every new task as if it had no relevant history.
How ReasoningBank structures memory
ReasoningBank organizes distilled knowledge into structured memory items, each composed of three fields:
- Title: a short identifier capturing the core strategy
- Description: a brief summary of the memory entry
- Content: the distilled reasoning steps, decision rationales, or operational insights pulled from past experiences
The system operates in a continuous loop. Before acting, the agent retrieves relevant memories from the bank and incorporates them into its context. After completing or failing a task, a language model acting as a judge evaluates the resulting trajectory. The agent then extracts either success insights or failure reflections and consolidates them back into the bank.
The researchers note that the self-judgment step does not need to be perfectly accurate. They found ReasoningBank remains robust even when those assessments contain noise — a practical consideration for real-world deployment where ground-truth evaluation may not always be available.
What the benchmarks show
Google evaluated ReasoningBank on web browsing and software engineering benchmarks. Compared to baseline approaches, the framework improved both agent effectiveness, measured as higher task success rates, and efficiency, measured as fewer steps required to complete tasks.
The combination of those two gains matters in practice. An agent that succeeds more often while also taking fewer actions is both more reliable and cheaper to operate at scale.
Why this approach differs from fine-tuning
ReasoningBank is explicitly a test-time mechanism. It does not retrain the underlying model; instead, it augments the agent's context with accumulated strategic knowledge at inference time. This means the memory system can evolve continuously after deployment without requiring retraining cycles, which are costly and slow for production systems.
The distinction between storing raw trajectories and storing distilled reasoning is central to the framework's design. A trajectory log tells an agent what happened; a ReasoningBank entry is intended to tell it why a decision worked or failed and how to generalize from that outcome.
Availability
The paper, "ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory," was presented at ICLR. Google has released the framework code publicly on GitHub alongside the publication.