Accurate identification of individual animals, known as animal re-identification (ReID), is crucial for long-term monitoring of behavior. Temporal tracking can only work well if animals are never occluded or never occlude each other. Thus, tools like DeepLabCut have built in methods for ReID (Lauer et al. Nature Methods 2022). For non-patterned species, such as mice, gorillas and fruit flies, ReID is particularly challenging due to the lack of salient body markings across individuals. Here we focus on one of the most difficult species: Brown bears (Ursus arctos). ReIDing this species is particularly challenging, as their appearance dramatically changes over a short period of time due to factors such as fur shedding, substantial weight gain, and scars from fighting. Addressing this, we have compiled an extensive longitudinal dataset of brown bears in Alaska (~70k images) to advance non-invasive ReID methods for unmarked species spanning multiple years (2017-2022). We propose and benchmark a novel, markerless ReID framework, called PoseReIDSwin. Our method outperforms other state-of-the-art methods from person ReID literature in different metrics. Furthermore, our model exhibits remarkable ReID performance across years, demonstrating the ability to identify bear images captured in a novel season, as well as promising results for identifying “novel bears,” or bears who have not been seen in previous years. PoseReIDSwin shows the potential to integrate behavioral analysis with ReID technology in a non-invasive manner, enabling further understanding of the individual or social behaviors of animals over time without the use of physical tagging or marking.