VentureBeat
Follow
New memory framework builds AI agents that can handle the real world's unpredictability
ReasoningBank is a new framework that empowers large language model (LLM) agents to learn and improve by organizing experiences into a memory bank. This framework distills generalizable reasoning strategies from both successful and unsuccessful attempts at solving problems. The agent uses this memory during inference, avoiding past mistakes and making better decisions on new tasks. ReasoningBank outperforms traditional memory mechanisms across web browsing and software engineering benchmarks. Current LLM agents often fail to learn from accumulated experiences, repeating mistakes and missing valuable insights. ReasoningBank addresses this by transforming each task experience into reusable reasoning memory. This allows agents to adapt proven strategies from past similar cases. The framework utilizes LLM-as-a-judge scheme to judge success and failure eliminating the need for human labeling. Agents retrieve relevant memories to guide actions by incorporating the memory into their reasoning process. Memory-aware Test-Time Scaling (MaTTS) enhances performance by integrating scaling with ReasoningBank. MaTTS incorporates parallel & sequential scaling, further boosting performance. ReasoningBank showed improved success rates and reduced interaction steps in web browsing tasks. It has a direct impact on operational costs, especially in eliminating trial-and-error. ReasoningBank offers a practical pathway towards building adaptive and lifelong-learning agents for enterprises and applications.
