A new method for updating the template in a feature tracking application is presented, which has minimal memory and processing overhead. The proposed method is an expectation maximisation inspired approach based on modelling the variable appearance of a template using a Gaussian mixture model in a discrete metric space, termed the M^3 I tracker for short. The proposed technique is compared to various other techniques in several experiments, where it performs robustly. Several comparison methods are outperformed. In addition to robust template tracking it has wider applications to advanced techniques such as AAMs and deformable templates.