A system for of optimizing delivery of multimodal content based on an adaptive, predictive caching algorithm. The system includes predictive caching software that receives at least one input regarding a preferred modality (e.g., audio, visual, or touch) for a particular application state. An application can have any number of states, each of which can, in turn, be presented in a different mode. The presentation mode for any particular application state can be selected by a user's input or request. In response to at least one input, the value of a weight variable can be modified or maintained. The value corresponds to a preferred presentation mode for an application state. The system can cache content for more efficient delivery in at least one presentation mode according to the weight variable.