Answering by approximate retrieval or by understanding reasoning are two ends of a spectrum.
Humans are at various places on this spectrum, depending on the task, experience, and depth of understanding.
We see this in physics or math students: some will study very hard, do lots of problems, learn solution templates, and get a passing grade. Others will barely study and get top grades. The difference? Mental models that enable reasoning.
The same is true for AI. Current LLMs are pretty close to the "retrieval" end of the spectrum. They don't have good mental models.
That's what we need to work on to get to the next level in AI: mental models that can be used for reasoning.
Unfortunately , too few people understand the distinction between memorization and understanding. It's not some lofty question like "does the system have an internal world model?", it's a very pragmatic behavior distinction: "is the system capable of broad generalization, or is it limited to local generalization?"