This paper deals with the evolution of embedded systems software at run-time. To accomplish such software evolution activities in resource-constrained embedded systems, we propose a component-based, execution time evolution infrastructure, that reconciles richness of evolution alternatives and performance requirements. The proposition is based on fine-grained optimization of embedded components, and on off-site component reifications called mirrors, which are representations of components that allow us to treat evolution concerns remotely and hence to reduce the memory footprint. An evaluation on a real-world evolution scenario shows the efficiency and relevance of our approach.