This paper proposes an agent-based framework using large-scale language models (LLMs) to address the challenges of documenting tacit knowledge within organizations. To address the challenges of incomplete initial information, difficulty in identifying experts, interaction between formal hierarchies and informal networks, and the need for appropriate questions, we simulate the knowledge dissemination process based on the Susceptible-Infectious (SI) model with decreasing propagation power. Through 864 simulations using various company structures and propagation parameters, the agent achieves a complete knowledge recall rate of 94.9%, and the self-critical feedback scores are strongly correlated with external literature evaluation scores. By analyzing the effects of each simulation parameter on the knowledge retrieval process, we show that information can be recovered without directly accessing domain experts. This highlights the ability of the agent to navigate organizational complexity and capture fragmented knowledge that is otherwise inaccessible.
