This paper introduces TradingGPT, an innovative multi-agent framework built on Large Language Models (LLMs) specifically designed for financial trading applications. The system addresses a key limitation of conventional LLMs—their inability to effectively prioritize immediate tasks due to non-hierarchical memory processing—by implementing a layered memory architecture with custom decay mechanisms that more closely emulate human cognitive processes. This hierarchical approach enables agents to better extract and utilize relevant insights from complex financial data, which is crucial for informed trading decisions in dynamic markets.

The framework features multiple agents that can engage in inter-agent debates about investment strategies, with each agent equipped with individualized trading traits to enhance memory diversity and decision robustness. By integrating multi-source historical actions and market insights through this layered memory system, TradingGPT demonstrates improved responsiveness to both past trades and real-time market signals. The authors assert that these sophisticated designs result in superior automated trading accuracy for stock and fund trading, positioning TradingGPT as a significant advancement in applying LLM technology to quantitative finance.