Recent advances in Reinforcement Learning (RL) have enabled agents to achieve remarkable performance in complex sequential decision problems. However, the opaque nature of RL models limits interpretability and generalization, prompting the need for hybrid neuro-symbolic approaches. In this paper, we propose an agent architecture that integrates reinforcement learning with a layered graph-based knowledge base (KB) offering a novel symbolic inference mechanism for inductive reasoning. The proposed architecture does not rely at all on knowledge provided by humans. The KB stores learned policies in a structured form, enabling explicit reasoning and generalization through a clique-based inference method applied to multigraphs with colored edges. We validate our approach on a grid-based environment, demonstrating improved learning efficiency and policy interpretability across deterministic and stochastic settings. Results indicate that agents leveraging the KB outperform purely sub-symbolic counterparts in cumulative reward and efficiency. Our findings suggest that combining reinforcement learning with structured symbolic reasoning offers a promising path toward explainable and adaptive decision-making systems.