Google Cloud AI's ReasoningBank: LLM Reasoning Decoded

For agents that think, memory is no longer a crutch, it’s a catalyst. ReasoningBank, powered by MaTTS, transforms the leftover debris of failed rollouts and successful paths into a structured, reusable intelligence. Parallel trajectories become contrastive lessons.

Sequential refinements become iterative insight. The result is a system that learns from its own reasoning, not after the fact, but as it goes. The numbers speak plainly: nearly nine percentage points higher success on WebArena, with fewer clicks.

That reduction in interaction steps is not a minor efficiency gain. It signals a deeper shift. The agent stops guessing and starts recalling.

It knows where past attempts stumbled and where they soared. What MaTTS proves is that scaling alone is blunt. Memory-aware scaling is sharp.

It turns compute into understanding, diversity into reliability. The positive feedback loop it creates is self-reinforcing: better memory begets better trajectories, which in turn beget even sharper memory. This is the kind of architecture that doesn’t just answer questions, it improves the way it approaches them.

And that is where the real breakthrough lives.