This paper introduces LD-Agent, a model-agnostic framework for long-term dialogue systems that integrates event memory, persona extraction, and response generation modules. The framework is designed to maintain long-term coherence and consistency in dialogue by leveraging historical event summaries and dynamic persona modeling. The event memory module uses long- and short-term memory banks to store and retrieve historical and ongoing session information, while a topic-based retrieval mechanism enhances memory accuracy. The persona module dynamically models user and agent personas, which are then integrated into the response generation module to produce appropriate responses. The framework is evaluated on two long-term dialogue datasets, MSC and CC, demonstrating its effectiveness, generality, and cross-domain capabilities. LD-Agent outperforms existing methods in terms of performance and shows strong adaptability across different models and tasks. The framework is also tested in cross-domain settings, where it performs competitively, nearly matching in-domain training results. The paper also presents ablation studies and human evaluations, highlighting the framework's ability to maintain coherence and consistency in long-term dialogue. The results show that LD-Agent is a powerful and generalizable framework for long-term open-domain dialogue systems.This paper introduces LD-Agent, a model-agnostic framework for long-term dialogue systems that integrates event memory, persona extraction, and response generation modules. The framework is designed to maintain long-term coherence and consistency in dialogue by leveraging historical event summaries and dynamic persona modeling. The event memory module uses long- and short-term memory banks to store and retrieve historical and ongoing session information, while a topic-based retrieval mechanism enhances memory accuracy. The persona module dynamically models user and agent personas, which are then integrated into the response generation module to produce appropriate responses. The framework is evaluated on two long-term dialogue datasets, MSC and CC, demonstrating its effectiveness, generality, and cross-domain capabilities. LD-Agent outperforms existing methods in terms of performance and shows strong adaptability across different models and tasks. The framework is also tested in cross-domain settings, where it performs competitively, nearly matching in-domain training results. The paper also presents ablation studies and human evaluations, highlighting the framework's ability to maintain coherence and consistency in long-term dialogue. The results show that LD-Agent is a powerful and generalizable framework for long-term open-domain dialogue systems.