Tangled Program Graphs (TPGs) are highly modular, hierarchical representations for genetic programming that are well-suited to multitask learningMultitask learning in temporal sequence prediction tasks such as control and time series forecastingTime series forecasting. In this work, we expand the simple scalar register machines traditionally used in TPGs toTangled program graphs include vector and matrix memory and operations. This helps TPGsTangled program graphs evolve versatile agents that are capable of solving partially observable control and forecasting problems simultaneously. A single agent can predict actions in discrete and continuous control tasks, as well as perform generativeTangled program graphs time series prediction.