There is an increasing variety of pervasive devices in use today. More and more applications are being supported on such devices. This requires device-specific application adaptation. We address the problem of speech application adaptation by dialog call-flow reorganisation for pervasive devices with different memory constraints. Given a dialog call-flow C and device memory size M, we present deterministic algorithms that alter C to create Cm that fits M by increasing the number of questions and splitting the underlying grammar. We can split a grammar by exposing the intermediate non-terminals in the grammar. The following observation forms the cornerstone of this paper: An and-grammar can be split "horizontally", and an or-grammar can be split "vertically" into its components to reduce the memory requirement of the call-flow, at the expense of increasing the number of prompts (questions). We present algorithm G-split, explain its implementation with example call-flows authored in VXML containing SRGS grammars.