Randomly choose 6 words from a 10k English dictionary and you're set. Hell, make it 12 words. Still insanely easier to remember than a 12 character alphanumeric/symbol password, and much more difficult to crack.
It has 50,105 words. Gets you an extra 2.3 bits per word.
Not exactly true if they are separate words. In a word-based password scheme, you are treating entire words, not characters, as units. The chance of fire, truck, and firetruck appearing in one password in that sequence, given a 10K word dictionary, is 1/(10K^3). The miniscule possibility of this speaks to the fact that there actually is a large entropy.