Open-domain dialogue systems, while exhibiting exceptional potential for wide-ranging application scenarios, suffer from hallucinations and weaknesses in maintaining long-standing personality consistency in multi-turn conversations. To mitigate this problem, the Localized and Generalized Persona Expansion (LoG-P) is proposed to enhance personalized dialogue system for consistent response generation. The LoG-P incorporates interaction learning and metamorphic relation construction as key components in localized and generalized persona expansions, aiming at enhancing persona understanding and consistency in response selection. Those make it a well-rounded solution for accurate personalized response selection in multi-turn dialogue generation. Since the retrieval-augmented generation (RAG) can enhance generative dialogue systems by retrieving relevant responses from available data stores, leading to higher retrieval accuracy, we first conduct the experiments on two assigned persona versions focusing on consistent personality reveal that the LoG-P significantly improves retrieval accuracy over the state-of-the-art methods.