Not really. DIMM modules are fairly cheap but DIMM slots are not. Your average mid-range server can, in theory, be maxed out to 192G or even 256G RAM. But only if you use 8G or even 16G DIMMs - and those are still hideously expensive.
The most cost efficient configuration usually gravitates around 64G (16x 4G) nowadays, rarely higher.
So, when you factor in the cost overhead per server (rackspace, power/cooling, network infrastructure, maintenance) then it's not hard to understand why RAM is the most expensive ressource in the cloud.
Also, with deploy hooks at EY, I have command line deployment similar to Heroku's "git push heroku master" by using a specific command in your commit message 'git commit -am "[deploy myapp_production]" and then 'git push'.
In addition, there is absolutely nothing stopping you from running your own, separate memcached server and pointing your Heroku app there. If you host it on EC2, then you should have very low latency to Heroku.
Heroku is obviously not perfect, but this particular criticism is not valid. That said, you can point this criticism at Northscale.
Heroku is indeed great for starting out when you can get away with the free plan or need only a small number of dynos and ideally none of their addons. Heroku can be a beautiful launchpad during the bootstrap phase.
Once your heroku bill approaches about ~$500/month you should start looking elsewhere, though. Their pricing for larger deployments becomes outright hilarious above roughly that threshold.
And with hilarious I mean really hilarious, as in the $3500 price point for a 50G memcached instance. For that money you can also buy a physical server with 64G RAM every month, fresh from the factory...