The recent advances in the chronic implantation of electrodes have allowed the collection of extracellular activity from neurons over long periods of time. To fully take advantage of these recordings, it is necessary to track single neurons continuously, particularly when their associated waveform changes over time. Multiple spike sorting algorithms can track drifting neurons but they do not perform well in conditions like a temporary increase in the noise level and changes in the number of detectable neurons. In this work, we present Tracking_Graph, a general framework to track neurons under these conditions. Tracking_Graph can be implemented with different spike sorting algorithms, allowing the experimenter to use the algorithm best fitted for their setup. The main idea behind Tracking_Graph is the blockwise analysis of the recording and application of a classification algorithm to match spikes to templates in different segments, leading to a directional metric that can be used to link clusters across blocks. Moreover, the algorithm can detect and fix sorting errors (splits and merges) in isolated blocks. We compared an implementation of Tracking_Graph with other algorithms using long-term simulations and obtained superior performance in all the metrics.