ReasoningBank: Giving AI Agents a Memory That Actually Learns from Failure

Google Research just dropped something interesting at ICLR this year: ReasoningBank. It’s a memory framework for AI agents that doesn’t just log what happened — it actually learns from both wins and screw-ups.

If you’ve been following the agent space, you know the pain. Agents are great at navigating web pages or hacking through codebases, but once they’re deployed and running long-term, they hit a wall. They don’t learn from experience. They’ll make the same dumb mistake a hundred times because they have no mechanism to reflect on what went wrong.

Existing memory approaches are either too granular or too narrow. Some save every single action as a trajectory — think Synapse’s approach. Others only document workflows from successful runs, like Agent Workflow Memory. Both miss the point. Saving every click doesn’t teach strategy. Ignoring failures means throwing away the most valuable data you have.

ReasoningBank takes a different angle. Instead of raw action logs, it distills high-level reasoning patterns. Each memory item has a title, a description, and the actual distilled content — the “why” behind the steps. The agent retrieves relevant memories before acting, then self-assesses the outcome using an LLM-as-a-judge. If it works, great, extract the insight. If it fails, even better — extract the lesson.

What I like about this is the focus on failure. Most systems treat failures as noise to be ignored. ReasoningBank actively mines them for counterfactual signals. Instead of learning “click the ‘Load More’ button”, it learns “first check the page identifier to avoid infinite scroll traps”. That’s a tactical insight, not a procedural rule. It transfers across tasks.

The self-judgment doesn’t need to be perfect either. The authors found the framework robust against noisy judgments, which is reassuring because LLM-as-a-judge is famously unreliable in edge cases. They also keep things simple — new memories just get appended, no fancy consolidation. That’s a pragmatic choice for now.

On benchmarks for web browsing and software engineering, ReasoningBank improved both success rates and efficiency — fewer steps to complete tasks. That’s the kind of win that matters in production.

Is this the final answer to agent memory? Probably not. The consolidation strategy is still naive, and long-term memory retrieval at scale is its own can of worms. But it’s a solid step toward agents that actually get better over time instead of plateauing after deployment.

If you’re building persistent agents, this is worth a close look. The code is on GitHub, and the paper is at ICLR 2026.