ReasoningBank: Agent Memory Shifts from Saving Trajectories to Distilling Transferable Strategies from Successes and Failures
Google Research introduced the ICLR paper ReasoningBank in April 2026, which enables agents to form structured reasoning strategies from success and failure trajectories through a closed loop of retrieval, experience distillation, and memory consolidation.
Existing agent memory often saves facts or complete action trajectories but fails to convert failures into executable strategies for the next time. ReasoningBank compresses experience into cross-task reusable reasoning rules, making test-time learning a new capability layer for persistent agents.
Each memory contains a title, description, and reasoning content; before execution, the agent retrieves strategies, after completion it self-evaluates the trajectory and distills successful experiences or failure reflections, then merges new rules back into the memory bank. The paper reports both success rate improvements and task step reductions on web browsing and software engineering benchmarks, and releases code publicly.
Competition among long-running agents will shift from context length to experience governance: which failures are worth remembering, how to merge conflicting rules, and when to retire outdated experiences will affect real-world task cost and stability.
When purchasing persistent agents, demand demonstrations of cross-week task reuse, failure non-repetition rate, memory retirement, and audit capabilities, rather than just claiming support for long-term memory.
It is necessary to verify whether self-evaluation errors and malicious trajectories can contaminate memory, and to compare long-term benefits, memory maintenance costs, and cross-model transferability in real enterprise tasks.