I maintain an offensively big knowledge graph, and I feel like you're both right -- building those connections manually is both mind-expanding and somewhat labor intensive.
If I have a conversation with X and discover they're an expert about Y, learning fact Z, I like to record Z as a file, and link it to both X and Y. That forces me to keep X and Y as entities in the graph. Which, if I care about them, is worth doing.
The real beauty of making those connections is the act of classifying something can cause you to explicitly recognize what concepts it is related to, and formulate precisely what the connection is.
The real pain point is knowing what your future self will wish your past self kept a file on. Some ideas (generally more social science-related ones, but also something like a bug involving the interaction of a lot of disparate technologies) are so cross-cutting that I'd need to make five parent concepts to file them "completely", and often then I'd like to file those parent concepts further. I generally don't.