The document presents a novel adaptive myoelectric decoding algorithm to improve long-term classification of electromyography (EMG) signals for prosthesis control. The algorithm relies on unsupervised updates to the training set to adapt to slow and fast changes in EMG signals over time. It was tested over 4.5 hours with 16 subjects performing 8 hand gestures. The adaptive algorithm showed significantly less decay in decoding accuracy compared to a non-adaptive system, maintaining over 86% accuracy on average over the full testing period.