Parametric memory · Model knowledge
Parametric memory is the knowledge encoded in a model’s weights during training, as opposed to information retrieved live from the web. It never produces a citation: the model answers without being able to point to a source page.
Standalone definition · 38 words
How does parametric memory work?
During training, the model compresses large amounts of text into its parameters. At generation time, it reproduces what it retained, with the inaccuracies and knowledge cutoff of its data. The term comes from RAG vocabulary: the founding paper by Lewis et al. (2020) distinguishes parametric memory, inside the model, from non-parametric memory, in the document index.
Why does the distinction matter for GEO?
An answer drawn from parametric memory may name or omit a brand with no link or source. An answer drawn from grounding may cite it with a link. The levers differ: the first depends on the brand’s presence in training data, the second on how retrievable the pages are.
Can you influence parametric memory?
Indirectly and slowly: by being mentioned consistently across many reliable sources before the next training runs. There is no direct lever, and no control over updates. To measure it, query models without web access.
Frequently asked questions
Does blocking GPTBot remove the brand from the model’s memory?
It prevents future collection by that bot, not the erasure of what has already been learned.
Is parametric memory up to date?
No, it stops at the training data cutoff. That is why engines add web search for current events.
Can a model hallucinate about a brand?
Yes, especially if the brand is poorly documented. Grounding reduces this risk without eliminating it.
Related terms
Further reading
- Checklist GEO and LLM / AI 2026
- 2026 GEO Tools Comparison: The Decision Guide
- ← Back to the GEO glossary (114 terms)