The AI can also give you pretty good examples of "kind" that you can then evaluate. I've had it find companies that "do X" and then used those companies to understand enough about what I am or am not looking for to research it myself using a search engine. The last time I did this I didn't end up surfacing any of what the AI provided. It's more like talking to the guy in the next cubicle, hearing some suggestions from them, and using those suggestions to form my own opinion about what's important and digging in on that. You do still have to do the work of forming an opinion. The ML model is just much better at recognizing relationships between different words and between features of a category of statements, and in my case they were statements that companies in a particular field tended to make on their websites.

Live performance (like conversation or playing music) often relies on memory to do it well.

That might be a good criteria for how much to memorize: do you want to be able to do it live?

Pilots have both checklists that they can follow without memorizing, but also memory items that have to be performed almost instinctively if they encounter the precondition events.