How to incorporate last agent actions as input to next action decision

[kailuowang]

First of all, is this actually needed?
In deepmind's use cases the screen capture all information, in our case can our input capture all information?
If so how to make sure that we don't lose any information if action input is discarded.
If not then how to incorporate the action back into the readings ( since it's not a one-to-one match with the reading).