"my wife" is Almost Certainly* explicitly flagged in his address book, either as an alias or through a "relationship" field. he didn't show you that part. (the same is likely true of the geofence, though a location-aware algorithm could probably do a better job inferring where your work is than who your wife is.)
I have no problem with that. This a great demonstration of what Apple does best - optimizing products to fit the needs of most users. Google already has the voice tech in place, but often fails to follow through on the usability front. A perennially annoying example for me is when I'm on the road somewhere, bring up Maps, and ask for directions from wherever-I-am to 'home'; it doesn't know where 'home' is. This is even more annoying, because there is a 'home' tag in Maps, but it doesn't get searched on. Likewise, there's no easy way to tell it that one person is my wife, another person is my business partner, a third person is my nemesis and so on.
Apple are leveraging a culture of paying attention to detail to overcome technical handicaps, but it's hard to quantify or A/B test that, and most firms aren't comfortable with a budget line item called 'magic.'
I wonder how many conditionals you can string together. "Remind me to call my wife before I leave work" is great, but "remind me to call my wife before I leave work at the end of the day, unless she calls me first" would be awesome.
In reality none of this stuff is even remotely useful, because I can't even begin to guess whether or not a given command will be interpreted properly. Natural language is hard, and reminders need to be 100% reliable.