A method for "preempting memory" is presented, where (parts of) the memory allocated to an active task may be reallocated to another task, without corrupting the state of the active task's job. The method is based on combining scalable components with Fixed-Priority Scheduling with Deferred Preemption (FPDS). Real-time systems composed of scalable components are investigated. A scalable component can operate in one of several modes, where each mode defines a certain trade off between the resource requirements and output quality. The focus of this paper is on memory constrained systems, with modes limited to memory requirements. During runtime the system may decide to reallocate the resources between the components, resulting in a mode change. The latency of a mode change should satisfy timing constraints expressed by upper bounds. A modeling framework is presented combining scalable components with FPDS. A quantitive analysis comparing Fixed-Priority Preemptive Scheduling (FPPS) and FPDS is provided, showing that FPDS sets a lower bound on the mode change latency. The analytical results are verified by simulation. The results for both FPPS and FPDS are applied to improve the existing latency bound for mode changes in the processor domain. The presented protocol is especially suited for pipelined applications, allowing to perform the mode change without the need to first clear the whole pipeline.