Parametric memory is the knowledge a large language model absorbed during training and stored inside its parameters (its weights). When a model answers from parametric memory, it is drawing on what it learned, not on any document it fetched for that specific question. It is the model's built-in, from-memory knowledge.
This is the counterpart to retrieval. In retrieval-augmented generation, an engine fetches live documents and grounds its answer in them; without retrieval, the model falls back on parametric memory alone. The two behave very differently for a brand. Parametric memory is frozen at the knowledge cutoff and changes only when the model is retrained, so it can describe an outdated version of you with complete confidence. Retrieved knowledge is as current as the page just read.
For AI visibility, the distinction sets your priorities. You influence parametric memory only slowly and indirectly, by being consistently and correctly described across the web so the next training cycle learns the accurate version. You influence retrieval directly and now, by being reachable and citable. An answer with no sources is usually parametric memory talking; an answer that cites pages retrieved them.