This paper describes a new memetic semantic algorithm for symbolic regression (SR). While memetic computation offers a way to encode domain knowledge into a population-based process, semantic-based algorithms allow one to improve them locally to achieve a desired output. Hence, combining memetic and semantic enables us to (a) enhance the exploration and exploitation features of genetic programming (GP) and (b) discover short symbolic expressions that are easy to understand and interpret without losing the expressivity characteristics of symbolic regression. Experimental results show that our proposed memetic semantic algorithm can outperform traditional evolutionary and non-evolutionary methods on several real-world symbolic regression problems, paving a new direction to handle both the bloating and generalization endeavors of genetic programming.