<p>Personalized recommendation systems are vital in modern online services, enhancing user experiences by tailoring suggestions to individual preferences. Recently, the integration of Knowledge Graphs (KGs) has gained traction for improving recommendation quality and transparency by leveraging rich semantic structures. However, existing KG-enhanced methods largely emphasize items aligned with user preferences, overlooking the insights from negatively associated items. Therefore, we propose DualPMPR (Dual Policy-guided Multi-hop Path Reasoning), a novel dual-agent framework that reasons over multi-hop paths within a KG under a unified reinforcement learning paradigm. The positive agent explores user-favored items, while the negative agent identifies counterfactual items that reflect user dislikes, together providing complementary preference signals. This design explicitly integrates negative preference reasoning and causal explanations into the recommendation process, advancing beyond traditional KG-enhanced models. From an engineering perspective, we apply DualPMPR to recommendation tasks in domains such as e-commerce and digital content services. Experiments on five real-world datasets demonstrate that it consistently outperforms strong baselines in terms of precision, recall, and NDCG (Normalized Discounted Cumulative Gain). Moreover, the framework produces interpretable recommendation paths and counterfactual explanations, making it a promising solution for transparent and practical recommendation systems.</p>