Timely medical consultations and health management can significantly reduce the incidence of geriatric diseases and the risk of complications. Intelligent medical consultations allow patients to receive medical advice and health recommendations without leaving their homes, thereby alleviating the strain on medical resources and meeting the need for autonomous health management. However, the existing question-answering systems based on knowledge bases and knowledge graphs often perform poorly in complex multi-turn dialogues and dynamic environments due to a lack of flexibility and contextual understanding. The advent of Large Language Models (LLMs) provides the potential for constructing more intelligent and efficient medical consultation systems.However, some issues such as weak interpretability and Q&A hallucination remain. Thus a multi-agent architecture based on LLMs and knowledge graph technology is introduced to propose a more optimized solution for medical consultations on geriatric diseases, and an intelligent medical consultation system is developed based on it. The system employs an agent framework to distribute medical decision-making across multiple agents, each focusing on specific types of problems. By utilizing clear decision-making and execution strategies, the agents progressively search for optimal solutions, ensuring that the question-and-answer process is highly interpretable and transparent. The Geriatrics-MedLLM, a LLM for answering questions on geriatric diseases, is developed through pretraining and LoRA fine-tuning. Additionally, the system utilizes the SummaryBufferMemory to summarize key content after each round of dialogue, allowing for dynamic adjustments and optimizations of the dialogue content to ensure coherence and accuracy. Furthermore, a geriatric medicine dataset has been collected with web crawler technology from well-known Chinese medical information websites for two sets of quantitative experiments, simulating the Q&A process of 100 patients. The experimental results show that the system can provide more accurate medical advice in multi-turn dialogues with highly accurate and interpretable responses.