Mondrian Forests are a powerful data stream classification method, but their\nlarge memory footprint makes them ill-suited for low-resource platforms such as\nconnected objects. We explored using reduced-precision floating-point\nrepresentations to lower memory consumption and evaluated its effect on\nclassification performance. We applied the Mondrian Forest implementation\nprovided by OrpailleCC, a C++ collection of data stream algorithms, to two\ncanonical datasets in human activity recognition: Recofit and Banos \\emph{et\nal}. Results show that the precision of floating-point values used by tree\nnodes can be reduced from 64 bits to 8 bits with no significant difference in\nF1 score. In some cases, reduced precision was shown to improve classification\nperformance, presumably due to its regularization effect. We conclude that\nnumerical precision is a relevant hyperparameter in the Mondrian Forest, and\nthat commonly-used double precision values may not be necessary for optimal\nperformance. Future work will evaluate the generalizability of these findings\nto other data stream classifiers.\n