Listen: Parametric Memory
Knowledge held in a model's own weights and recalled at inference with no lookup, as distinct from non-parametric memory, which is text retrieved into the context window at request time.
Transcript
When an artificial intelligence model answers a question instantly, without searching the web or checking a database, it is relying on its parametric memory. This is the knowledge baked directly into the model's weights during its training phase. It is different from non-parametric memory, which is the external information a system retrieves and feeds into the context window at the moment you ask a question.
Inside the model, facts are stored across layers of connections that act like key-value memories. How much a model can remember depends heavily on exposure. If a fact appears thousands of times in the training data, it is encoded strongly. If it only appears a few times, the model might not remember it at all.
Because this memory is frozen when training ends, it cannot keep up with real-time events. Even worse, when a model is asked about a weakly stored fact, it will still generate a fluent, confident answer. The weights carry no separate signal to indicate a fact was never stored, which is the main cause of artificial intelligence hallucinations.
This division shapes how information appears in AI answers. Well-known subjects are recalled directly from parametric memory, often with no links or citations attached. Less familiar subjects only exist for the model when they can be retrieved from the live web. This puts a heavy premium on crawlability and freshness, as retrieved content is what allows the model to cite its sources.