If your team builds massively parallel systems in Erlang, you need to make sure a candidate understands at least the basics of its process model and message passing. If your team build high-performance web apps, you need to make sure a candidate understands at least the basics of HTTP and the difference between client and server. For SQL, the same is true: they need to understand at least the basics of the relational model and declarative programming.
This is not complex problem solving, this is problem solving. I'm not sure what memorization has to do with any of this. SQL is a language with only maybe 10-15 important words. It's not like trying to get around Paris without a phrasebook as a non-French speaker.
I've found that using graphical query writers (I only know of the one in access) and ORMs are generally harder than writing the SQL. ORMs are good for keeping code db agnostic, though.
Being able to write SQL queries from memory has little correlation to a candidate's level of ability. Personally I consider myself a fairly strong developer and it hasn't been only until the last year that I can now write pretty complex joins from memory. And I've been developing for 20 years. Only because of a recent project and the volume of queries I had to write did my method change from using a graphical query writer to simply memorizing the syntax I need. Indeed, this very type of adaptation is something I look for in candidates.