Google Research Releases ReasoningBank: Agent Memory from Failures
Google Research has released ReasoningBank — an agent memory system that separates how it learns from success and failure trajectories. Prior systems (Synapse, AWM) stored only successes; adding failures to AWM caused a -2.2% accuracy drop. ReasoningBank addresses this by extracting validated strategies from successes and discrete lessons from failures, then abstracting both away from specific website contexts so strategies generalize. Results: +8.3pp WebArena success rate with Gemini-2.5-flash, 54%→57.4% on SWE-Bench-Verified with +4.3% token overhead. MaTTS (Memory-aware Test-Time Scaling) builds a self-reinforcing loop on top.
Why It Matters
Treating failure trajectories as first-class learning signal — not noise to discard — is a structural advance in agent memory design with direct implications for long-running agentic systems.