Parametric memory
Parametric memory is what an AI model learned in training and holds in its own weights, as opposed to what it looks up when it answers.
Parametric memory is what an AI model knows from its training, stored in the model itself. When a model answers without searching the web, everything it says comes from this memory. The term comes from the research that introduced RAG, which described a model's own knowledge as its parametric memory and the documents it looks up as its non-parametric memory.
It matters to marketers because it's slow to change. What a model learned about you in training stays with it until the next version is trained, even if you've changed your prices, your products or your name since.
How it shows up in answers
Ask a platform about your brand with web search off, or on a platform that decides not to search, and you're reading its parametric memory. That's where old facts live.
For example, say Harrow & Pine stopped making pine furniture two years ago and now sells only oak. A model trained on pages from before the change may still describe Harrow & Pine as "a maker of affordable pine furniture." Updating the website fixes what the model reads when it searches. It does nothing to what the model already remembers.
What you can do about it
You can't edit a model's memory directly. You can do two things.
First, make sure the platforms find the right page when they search. Some platforms search the web for every answer, and others decide for themselves. When they do search, what the model reads can outweigh what it remembers.
Second, make the correct facts common on the web: your own site, directories, review sites, press. The next model is trained on what's out there, so the more often the right version appears, the better its chance of sticking.
In AnswerOrbit
When an answer says something wrong about you without using any page, Inaccuracy drivers in Claims shows the platform's own knowledge as the cause. From the claim, Report to and the platform's name asks for an evidence pack you can send to the platform.