This paper introduces a neuro-symbolic framework designed to predict and explain subsequent facts from current observations. Facts are generated through causal relationships, which can be modeled by a set of propositional logic rules representing the domain knowledge. However, these rules remain unknown to the agent. By observing the facts, the agent constructs an approximation of them, which is then used to predict and explain new facts. The proposed framework can learn and adapt to different environments modeled by various forms of logic programs, also handling negation and recursion. Most notably, it can handle dynamic environments whose structure evolves over time. In these scenarios, the agent modifies its understanding of the environment to capture new observations, guaranteeing that its model of the domain knowledge remains up-to-date. To achieve this goal, our approach leverages the A2C (Advantage Actor-Critic) reinforcement learning algorithm. This choice allows us to integrate reinforcement learning principles into our logic framework. Through this research, we aspire to contribute to the development of explainable neuro-symbolic Artificial Intelligence systems in dynamic environments.