Title: FundaPod: Leveraging Knowledge Graph Memory for Multi-Persona AI Agents in Fundamental Investment Research

Abstract:

While large language models (LLMs) are gaining traction in the financial sector, current applications predominantly focus on predictive financial NLP tasks or generating trading signals. In contrast, institutional fundamental research demands a different approach: human analysts or AI agents must collect evidence, pinpoint business drivers, weigh opposing perspectives, and draft investment memos. The primary objective here is not simply to forecast outcomes, but to create transparent, reusable, and verifiable investment plans that facilitate the cumulative growth of investment knowledge.

To address this, we introduce FundaPod, a platform designed for AI-assisted fundamental investment research utilizing multi-persona agents. We posit that fundamental research is a human-centric decision-support activity, qualitatively distinct from the generation of trading signals, and thus requires an architectural design that preserves agent independence. Within FundaPod, AI agents adopting distinct personas—such as macro strategists or value investors—conduct their research independently, bound by a shared provenance contract. Subsequently, a knowledge-graph memory system surfaces these agents’ divergent views for the human portfolio manager (PM) to adjudicate.

This work outlines five design principles for human-AI hybrid systems that support fundamental research, drawing from design-science methodologies and theories regarding cognitive isolation and human-machine coordination. Furthermore, we detail four core architectural mechanisms: a persona distillation pipeline that converts public investor materials into functional agents; a declarative skill registry enabling the planner to generate typed task graphs; a grounded evidence model that ties memo assertions to verifiable sources; and a knowledge-graph "second brain" that interconnects tickers, memos, analysts, and themes. We validate this architecture through a comprehensive case study and an analysis of persona-based memo comparisons.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC