This paper introduces a novel cognitive multipath multi-channel routing protocol (CMMRP) for mobile ad-hoc networks. It is designed for a multi-channel environment where nodes can simultaneously use multiple interfaces to transmit packets over different frequencies. It employs cognitive functions to make nodes intelligently select multiple node-disjoint, edge-disjoint and frequency- disjoint paths. Neural network machine learning is adopted to make nodes aware of history. CMMRP employs a modified path discovery protocol which can be divided into two parts, a space discovery and spectrum discovery. Simulation results show that CMMRP significantly improves network reliability and performance.