Large Language Models(LLMs)-driven Artificial Intelligence (AI) Agent provide conversational generation AI with human-like decision making and reasoning capabilities and extend them to Retrieval Augmented Generation (RAG) external knowledge base tools. However, the existing research focuses on knowledge coverage, ignores the problem of redundant knowledge accumulation, and fails to take into account the interpretability of LLMs, which limits the application value. In this paper, a dynamic and stepwise enhanced KNN framework, DKCER-Agent, is proposed. The framework effectively reduces redundant information through multi-round dynamic reasoning and gradually queries and generates optimal answers. The redundant information is filtered by the joint retrieval mechanism, and the key information is cleared and filtered by KNN retrieval. In addition, the framework introduces a step-by-step thinking revision mechanism to decompose the problem into sub-problems, which are verified and generated by the LLMs generation strategy, thus enhancing the interpretability of the model. The experimental results show that DKCER-Agent significantly exceeds the baseline model in a wide range of multi-round dialogue dataset tests, and the accuracy is significantly improved.