Knowledge neurons are specific neurons in a transformer's feed-forward layers that appear to store factual associations, identified through causal tracing or activation patching. Understanding and locating these neurons enables targeted model editing — updating or erasing specific facts without retraining the entire model.