In order to provide personalized interactions in a conversational system, responses must be consistent with the user and agent persona while still being relevant to the context of the conversation. Existing personalized conversational systems increase the consistency of the generated response by leveraging persona descriptions, which sometimes tend to generate irrelevant responses to the context. To solve this problems, we propose to extend the persona-agnostic meta-learning (PAML) framework by adding knowledge from ConceptNet knowledge graph with multi-hop attention mechanism. Knowledge is a concept in a triple form that helps in conversational flow. The multi-hop attention mechanism helps select the most appropriate triples with respect to the conversational context and persona description, as not all triples are beneficial for generating responses. The Meta-Learning (PAML) framework allows quick adaptation to different personas by utilizing only a few dialogue samples from the same user. Our experiments on the Persona-Chat dataset show that our method outperforms in terms of persona-adaptability, resulting in more persona-consistent responses, as evidenced by the entailment (Entl) score in the automatic evaluation and the consistency (Con) score in human evaluation.