To understand the neural basis of many cognitive processes, such as learning and memory, we need to track the millisecond, single-spike activity of neurons across days. However, compared to calcium imaging methods, tracking neurons in chronic electrophysiological recordings remains challenging. Recent studies have designed algorithms to do so, but mostly rely on engineered parameters that may not take advantages of the full complexity of the neuron’s electrophysiological waveform and miss or falsely match neurons. To go beyond these limitations, we developed DeepMatch, a deep neural network framework based on contrastive learning, to track neurons across days. In short, the network is optimized to maximize the similarity between waveforms from the same neuron and minimize similarity between those from different neurons. We evaluated DeepMatch’s performance on weeks-long Neuropixels recordings from the mouse brain. To get a close-to-ground-truth dataset, we first tracked pools of neurons across pairs of days using only their functional signatures, and not their waveforms. We used their firing rate correlations with other neurons, information not used by the tracking algorithms, and stringent selection criteria. We then compared DeepMatch’s ability to identify these functionally tracked neurons against UnitMatch, the current best method. DeepMatch outperformed UnitMatch by finding more functionally tracked neurons (more true positives), while both methods maintain a comparable number of total predictions. Additionally, DeepMatch generalized effectively to unseen recordings and unseen animals. These results establish DeepMatch as a deep learning-based tool for accurate, scalable and automated solution for long-term neural tracking.