This paper proposes a novel method called Memory-injected LLM Personalization (MiLP) to personalize large language model (LLM) responses by injecting parameterized user memory. The method leverages parameter-efficient fine-tuning (PEFT) and Bayesian optimization to efficiently incorporate user information into the LLM, enabling personalized response generation. The key idea is to parameterize user historical content and inject it into the LLM's architecture, mimicking real-world memory mechanisms. The approach involves using LoRA modules to store and activate user memory, and employing Bayesian optimization to search for optimal configurations for memory injection. The method is evaluated on three datasets, demonstrating significant improvements over existing baselines in terms of ROUGE-L scores and Persona-F1 scores. The results show that MiLP achieves superior performance in personalized response generation, with the effectiveness of the method being validated through both automatic and human evaluations. The study also highlights the importance of balancing the size of input information and trainable parameters to achieve optimal performance. The proposed method offers a promising approach for personalized LLM response generation, with potential applications in healthcare and other domains requiring tailored responses. The paper also discusses limitations, including the reliance on user historical content and the need for further research on scalability and efficiency. Overall, the study contributes to the field of LLM personalization by introducing a novel method that effectively incorporates user memory into the LLM's architecture.This paper proposes a novel method called Memory-injected LLM Personalization (MiLP) to personalize large language model (LLM) responses by injecting parameterized user memory. The method leverages parameter-efficient fine-tuning (PEFT) and Bayesian optimization to efficiently incorporate user information into the LLM, enabling personalized response generation. The key idea is to parameterize user historical content and inject it into the LLM's architecture, mimicking real-world memory mechanisms. The approach involves using LoRA modules to store and activate user memory, and employing Bayesian optimization to search for optimal configurations for memory injection. The method is evaluated on three datasets, demonstrating significant improvements over existing baselines in terms of ROUGE-L scores and Persona-F1 scores. The results show that MiLP achieves superior performance in personalized response generation, with the effectiveness of the method being validated through both automatic and human evaluations. The study also highlights the importance of balancing the size of input information and trainable parameters to achieve optimal performance. The proposed method offers a promising approach for personalized LLM response generation, with potential applications in healthcare and other domains requiring tailored responses. The paper also discusses limitations, including the reliance on user historical content and the need for further research on scalability and efficiency. Overall, the study contributes to the field of LLM personalization by introducing a novel method that effectively incorporates user memory into the LLM's architecture.