The topological structure of the search agents in the swarm is a key factor in diversifying the knowledge between the population and balancing the designs of the exploration and intensification stages. Marine Predator Algorithm (MPA) is a recently introduced algorithm that mimics the interaction between the prey and predator in ocean. MPA has a vital issue in its structure. This drawback related to the number of iterations that is divided into the algorithm phases, hence the agents do not have the adequate number of tries to discover the search landscape and exploit the optimal solutions. This situation affects the search process. Therefore, in this paper, the principle of the comprehensive learning strategy and memory perspective of the fractional calculus have been incorporated into MPA. They help to achieve an efficient sharing for the best knowledge and the historical experiences between the agents with the aim of escaping from the local solutions and avoiding the immature convergence. The developed fractional-order comprehensive learning MPA (FOCLMPA) has been examined with several multidimensional benchmarks from the CEC2017 and CEC2020 as challenging tested functions in the numerical validation part. For real-world applications, four engineering problems have been employed and a set of eighteen UCI datasets have been used to demonstrate the developed performance for feature selection optimization problem. The FOCLMPA has been compared with several well-regarded optimization algorithms via numerous statistical and non-parametric analyses to provide unbiased recommendation. The comparisons confirm the superiority and stability of FOCLMPA in handling the series of experiments with high qualified results and remarkable convergence curves.