FlowState is a proposed method for long-horizon LLM agents to retain and reuse historical information, improving performance and efficiency. It maintains semantic state nodes, relations, and references to raw tool observations, and enables agents to reassess prior decisions and guide subsequent actions.