Being based on human brain, neuromorphic computing systems are challenged by device degradation, environmental noise and scalability problems which reduce the reliability over time. In this paper, we present a novel self-healing architecture with bio-inspired astrocyte-like fault detection, memristor-based dynamic pathway reconfiguration and reinforcement learning to guarantee robust fault tolerance and recovery. By achieving 98.5% fault detection accuracy, 94% fault localization precision and 100% recovery success rate, yet complete signal integrity, our proposed methodology achieves fault detection, localization, and recovery both accurately and collaboratively. The system also can scale efficiently to networks with up to 10,000 neurons with minimal energy overhead (9.8%) and reconfiguration latency (5.8 ms). We validate the proposed architecture as a powerful and scalable solution for neuromorphic fault tolerant applications.