Good question Resition- If it is anything more than vaccum or a few atoms of gas then the drag on the nano-cantilevers would appear to have have the potential to stop the device from working. If the CNT fabric is spun on then it should be "relatively easy" to vacuum remove the solvent prior to metal electrode deposition.
It would be equally interesting to know in each NRAM memory cell how many nano-cantilevers are involved in each switching event. It would appear there is a large amount of redundancy so a few failures could be tolerated. In the 3D fabric a stuck on or stuck off would require opening up a few new paths. The cantilevers involved number might have implications for scaling.
There is a paper at SSDM 2013 that might have relevance to the performance of Nantera's CNT cantilever based memory. Titled "Release of Stuck Carbon Nanotube Arm by Resonant Vibration Toward Nano Relay Application, by T Tagota et al, as a joint effort by a team from Osaka Prefecture University and KRI Inc, Japan.
In the past Nantera reported that their memory had survived a trip in space, with all the vibration of blast off, it could be that test did not expose the memory to a critical resonant frequency. However, here again, as mentioned in my post below the redundacy afforded by the use of multiple cantilevers in each memory cell means that the loss of a couple by shaking may not be too much of a problem. It is unlikely they will all have the same resonant frequency.