Existing gait recognition frameworks retrieve an identity in the gallery based on the distance between a probe sample and the identities in the gallery. However, existing methods often neglect that the gallery may not contain identities corresponding to the probes, leading to recognition errors rather than raising an alarm. In this paper, we introduce a novel uncertainty-aware gait recognition method that models the uncertainty of identification based on learned evidence. Specifically, we treat our recognition model as an evidence collector to gather evidence from input samples and parameterize a Dirichlet distribution over the evidence. The Dirichlet distribution essentially represents the density of the probability assigned to the input samples. We utilize the distribution to evaluate the resultant uncertainty of each probe sample and then determine whether a probe has a counterpart in the gallery or not. To the best of our knowledge, our method is the first attempt to tackle gait recognition with uncertainty modelling. Moreover, our uncertain modeling significantly improves the robustness against out-of-distribution (OOD) queries. Extensive experiments demonstrate that our method achieves state-of-the-art performance on datasets with OOD queries, and can also generalize well to other identity-retrieval tasks. Importantly, our method outperforms the state-of-the-art by a large margin of 51.26% when the OOD query rate is around 50% on OUMVLP.