Finding low-cost designs of water distribution systems (WDSs) which satisfy appropriate levels of network performance within a manageable time is a complex problem of increasing importance. A novel multi-objective memetic algorithm (MA) is introduced as a solution method to this type of problem. The MA hybridises a robust genetic algorithm (GA) with a local improvement operator consisting of the classic Hooke and Jeeves direct search method and a cultural learning component. The performance of the MA and the GA on which it is based are compared in the solution of two benchmark WDS problems of increasing size and difficulty. Solutions that are superior to those reported previously in the literature were achieved. The MA is shown to outperform the GA in each case, indicating that this may be a useful tool in the solution of real-world WDS problems. The potential benefits from search space reduction are also demonstrated.