In theory, ANNs could have an output layer that passes data from one frame to another frame to assist things. But there's no real programming to "hardcode" something like object permanence into an ANN. You pretty much throw a bunch of data into the system and hope for the best.
Considering the path-planning requirements I would be absolutely shocked if Autopilot wasn't build history models and estimated paths for objects around the vehicle (other cars etc).
I expect what happened was that they trained their NNs for improved detection in one area but unknowingly reduced it in another. Perhaps now it can detect tricycles 99% but road barriers went down to only 30%. Having worked with NNs it's very common to see gains in one domain which come at a cost of reduced performance in another.
Barriers are a pretty big part of driving on roads and highways and the only reason it would have been unknowingly reduced would be if they just weren’t testing the NN against data with them.