When dealing with label noise, sample selection is a commonly used method in which data with small losses are typically considered to be correctly labeled. However, this method may fail to identify challenging examples with large losses, which are crucial for improving the model’s generalization performance. In this paper, we propose a novel selection strategy Noise-robust Learning via Full Consistency (NLFC) that measure the extent to which each example is forgotten during different training stages of the model. Based on this measure, we filter the examples and divide the data into clean and noisy sets. Furthermore, we address the issue of underutilization of labeled noisy examples by introducing consistency at the feature and image levels. Experimental results on benchmark-simulated noisy datasets and real-world noisy datasets demonstrate that our proposed method outperforms state-of-the-art methods.