Hierarchical Temporal Memory (HTM) is a neural network model that simulates the structure and function of the neocortex. To address the issues of high storage and training time overheads in the current HTM temporal memory, a heterogeneous graph-based HTM temporal memory algorithm is proposed. The algorithm decouples segments into pairwise neuron connections and utilizes a heterogeneous graph for storage. A multi-head collaborative predictive neuron search algorithm is designed to leverage multi-core or multi-processor systems, enhancing the training efficiency of HTM temporal memory. Experimental comparisons with existing HTM models and other sequence prediction methods demonstrate that our HTM improves prediction accuracy while reducing the time overhead of training by approximately compared to traditional HTM algorithms.