In this paper, a class of Clifford-valued stochastic neural network with delays is investigated, by using a direct method, that is, without decomposing the Clifford-valued system into a real-valued system. Based on the contraction mapping principle and contradiction, sufficient conditions are derived to ensure the existence and stability of almost automorphic solutions for the stochastic networks under consideration. Finally, two numerical examples are provided to show the feasibility of our results.