Information encoded in an LLM's weights and parameters during training, as opposed to retrieved external knowledge.