Every model I know about is retained from scratch on the raw data as often as possible. There are several reasons for this:
Anything derived from PII is also PII unless it is anonymized by aggregation. Everything that is PII is aggressively TTL'd from the time of its collection from the user. The privacy review team knows what they're doing and won't let you get away with clever lifetime extension tricks. If there is a model specifically for you, it isn't going to be causally connected to anything that's aged out.
You can delete individual items and categories from your history. It would be an unimaginable pain in the ass, both mathematically and engineering-wise, to propagate the updates for individual deletions through an incremental pipeline. Even worse, I can't imagine how you'd possibly argue that such an operation would satisfy a legal deletion request (HIPPA, GDPR).
If you need to change your learner - say you found a bug in your training code - but all you have is a trained network and no data, you're screwed. You lose all your history and your user's search, spam-filtering, or voice recognition quality tanks, and people really care about those things. Pure incremental pipelines are incredibly brittle.
All of these problems are trivially solved by retraining from the raw data. Disk is cheap. Compute is cheap. Engineer time is expensive. Fancy infrastructure is expensive. Mathematicians are super expensive. Legal compliance is fantastically expensive and grows in cost incredibly quickly with increased engineering complexity.
I work at Google. My views are not the views of my employer and so on and so forth. I build data quality infra for the knowledge graph. I haven't worked on a ton of projects that dealt with PII for public users, but even so the privacy reviews for the tools and pipelines I build get reviewed and occasionally changes are demanded to protect tool-use information about employees that use the infrastructure that I build.