FinAcumen Enhances Financial Reasoning with Self-Evolving Experience Memory

Financial multimodal reasoning tasks, which involve complex coordination of numerical computation, data retrieval, visual interpretation, and temporal analysis across various data sources, often challenge existing AI agents. While current tool-augmented agents improve execution, they typically lack persistent memory, leading to repeated rediscovery of strategies and failure patterns. This statelessness can result in unreliable tool routing, noisy data retrieval, and prone-to-hallucination reasoning in critical financial contexts. To address these limitations, researchers have introduced FinAcumen, a novel financial reasoning agent framework. This system is built around a selective experience memory that accumulates financially grounded reasoning experiences from previous interactions. It distills successful strategies and cautionary rules derived from failures into a permanent memory bank. During inference, FinAcumen selectively activates retrieved experiences only when their semantic relevance surpasses a predefined threshold, actively suppressing irrelevant memory through a fallback mechanism. The framework also incorporates a deterministic financial tool environment to ensure accurate numerical computation, retrieval, visual decoding, and answer verification. Benchmarking against four financial multimodal reasoning tasks, FinAcumen consistently outperformed a frozen 8B vision-language model, surpassed finance-specialized models, and rivaled leading proprietary general-purpose models, demonstrating improved reasoning reliability, especially under retrieval uncertainty.