if you have any understanding of how neural networks or other machine learning algorithms are implemented, you would know that attempting such a thing would be verging on madness, if not near impossibility.
I'm not convinced that 2k of RAM is unreasonable for machine learning. Ok, so maybe neural networks would be tricky, but RL? e-greedy doesn't take up that much RAM or program space, as long as your state space isn't huge. e-greedy is also relatively easy to implement.
This is something I've been meaning to play with, so I'm going to try it. If I get around to actually doing it, I'll post the code here.