World models for online learning in complex environments are increasingly essential, particularly in partially observable scenarios. Within this domain, biologically inspired models of temporal memory have emerged as a promising class of models. These models extend traditional artificial neural networks by incorporating the idea of dendritic segments into point neurons and leveraging the structural insights from neocortical layers to facilitate efficient online sequence learning. This paper introduces an innovative segment-growing strategy for enhancing the Distributed Hebbian Temporal Memory (DHTM) algorithm. The proposed modifications aim to boost the model’s generalization capabilities and introduce a more efficient segment growth approach based on real-time training statistics. Our findings demonstrate that this novel strategy not only reduces the number of segments used, thereby enhancing the model’s computational efficiency, but also enhances its predictive accuracy in challenging, noisy environments.