Precise factual understanding of molecules enables researchers to efficiently deduce molecular structures, properties, and prospective applications, thereby advancing practical tasks such as drug design and target identification. Traditional deep learning approaches, such as molecular representation learning, typically suffer from narrow applicability and limited generalization capabilities. Meanwhile, employing the text generation and content understanding abilities of large language models for molecular understanding is hindered by susceptibility to hallucinations. We propose that a multi-agent framework—where distinct language models are endowed with specialized roles—integrated with fact-driven principles for molecular understanding and inference, offers a promising strategy to alleviate the challenges outlined above. Drawing inspiration from contemporary jury mechanisms, we present MolJury, a role-based multi-agent system designed to achieve factual molecular understanding. We innovatively established role-defining rules for diverse agents in molecular understanding tasks, enhancing reasoning capabilities through multi-agent collaboration. We established evidence-driven molecular inference principles to systematically prevent hallucination phenomena. Experimental results on open-source datasets for molecular understanding tasks demonstrate the effectiveness of our proposed architecture and strategy.