Semantic memory is the general knowledge a language model contains as a result of training. It includes learned language patterns, factual associations, and domain knowledge present in the model parameters, but not information newly retrieved at inference time.
The information that an LLM has when training ends. For example, semantic memory includes an excellent knowledge of grammar, vocabulary, and the facts it was explicitly trained on. Semantic memory does not include information gathered from retrieval-augmented generation. Contrast semantic memory with episodic memory.