Parametric memory refers to knowledge that a neural network has learned from its training data and stored implicitly in its weights, or parameters. It is the knowledge a large language model draws on when answering without any external documents or tools, and is also called parametric knowledge.
The contrasting concept is non-parametric memory, in which material outside the model, such as a search engine or document store, is retrieved at inference time. Retrieval-augmented generation (RAG) is a representative method that combines the two. Because parametric memory cannot contain information after the training data cutoff and updating it requires retraining or fine-tuning, tasks that depend on current information typically pair it with external data sources.