It's not that difficult to memorize five random phrases, which will contain quite a bit of entropy(assuming the phrases are really selected reandomly) .
Except they stake-my-life-on-it will NOT be selected randomly if you allow any user to operate under that procedure. They will be selected "randomly" as in that same way that "DUCK SPATULA" is random: Not very.
If instead a cryptographically strong random number of, say, at least 128 bits is generated it can be directly converted into a secure mnemonic phrase with just 12 words (or fewer, if you want to deal with a bigger dictionary). This is probably way smaller than your "five random phrases" but by being ruthlessly specific about HOW its generated and what "random" means, it's really just another way of representing a 128 bit key, and it's perfectly fine to memorize that.
If your construction involves the human picking the words or phrases, as it does 99 out of 100 times someone says "brainwallet" then the actual entropy is much lower. Humans are poor sources of randomness, and telling them to be "random" actually seems to make them worse at it. Powerful statistical models are quite good at predicting what humans will do, and while they won't break every single instance they can break a surprisingly large number.
People who haven't worked on password cracking have this quaint notion of running a little dictionary file through a program... and this would have been accurate in 1990 for someone cracking at your unix-crypt uni shell account. Today the tools are significantly better and have been refined through the disclosure of hundreds of millions of unencrypted passwords and the same kind of statistical tools that power speech recognition and automatic human language transaction. This statistical intelligence gets backed up by the brute force of GPU and FPGA clusters that can try hundreds of million or even billions of attempts per second.
I'd rather not see cracking weak bitcoin brainwallets turn into a viable industry for people buying up no-longer-competative mining fpgas for pennies on the dollar.
Your brain is trained to memorize phrases, not random alphabet-number-symbol mix, it may indeed be the case that people can keep long series of english phrases in their memory longer than just some random strings, even if the length of the phrase series is much longer.
And whether it's secure to use really has to depend on how much money you will put in it, for pocket money it could be a viable alternative.
No, the security it depends more on how much all attackers eveywhere in the world are spending on setting up ultrafast hardware for brainwallet cracking. This number is likely to become very big: "just pocket money" has a way of growing through lazyness and the sum of all pocket money is great in any case. This is especially the case because the attackers can attack everyone with the same effort it would take to attack one person.
Viable alternative to what? Done right it is just a cryptographically strong number, so it's not really an alternative in that case, it just "is" the other thing.
Brainwallet is mainly designed to solve the deniability problem, so that people who could seize you disks can not use it as a proof against you, so it's a viable alternative to make a paperwallet and bury it under a oak tree. But yes, maybe it's not an alternative for storing pocket money.