ReasoningBank: Enabling agents to learn from experience
Google Research
Google Research built ReasoningBank, a memory system that lets AI agents learn from both their wins and their screwups. Most agent memory only saves what worked—this one also mines failures for lessons, and it actually boosts accuracy and speed on real benchmarks.
Based on reporting by Google Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Every AI agent has the same annoying flaw: it forgets. Ask it to navigate a website or fix a bug in a codebase, and if it fails today, it'll cheerfully make the exact same mistake tomorrow. Google Research's new paper, presented at ICLR, tackles this head-on with something called ReasoningBank, and the core idea is refreshingly simple — stop treating failure like garbage to be thrown away.
Most existing memory systems, like Synapse or Agent Workflow Memory, either log every single action an agent took (bloated and not very useful) or only save the recipe from tasks that succeeded. ReasoningBank does something different. It distills each experience — success or failure — into a compact memory item with a title, a short description, and the actual reasoning that led there. So instead of remembering "clicked Load More button," the agent might store something closer to a rule of thumb: check the page identifier before trying to load more results, because infinite scroll can trap you. That's the difference between a script and a lesson.
The system runs in a loop: retrieve relevant memories before acting, let an LLM judge how the attempt went, then extract and stash new insights for next time. Google's researchers found the judgment doesn't even need to be very accurate — the framework tolerates noisy self-assessment surprisingly well, which matters a lot for anyone worried about deploying this without a human checking every verdict.
They paired ReasoningBank with what they call memory-aware test-time scaling, or MaTTS, which lets an agent burn more compute per task — either running several attempts in parallel or refining one attempt step by step — and then mines all that extra exploration for better memories. Tested on Gemini-2.5-Flash across WebArena and SWE-Bench-Verified, ReasoningBank alone beat memory-free agents by 8.3% and 4.6% success rate respectively, while trimming almost three execution steps per task on the coding benchmark. Add MaTTS with five parallel runs and WebArena success climbs another 3%, with fewer steps still.
What's more interesting than the numbers is what the researchers watched happen over time: agents started with dumb checklist-style memories like "look for page links," and gradually built up compositional, preventative rules — the kind of thing a seasoned engineer might scribble in a postmortem. That's not just memory. That's something closer to accumulated judgment, and it's the part of this paper that should make people building long-running agents pay attention.
My take — AI-written commentary, not fact-checked reporting
This is the unglamorous infrastructure work that actually matters more than the next benchmark-topping model release, and I say that as someone who's tired of hype around agents that reset to zero every session. Learning from failure instead of just hoarding successful trajectories is such an obvious idea that it's a little embarrassing it took this long to formalize — but credit where due, Google shipped the code on GitHub instead of just teasing a paper, which is the right move if you actually want this adopted rather than admired.
Read more about this at: Google Research