For over a decade, the multi-agent patrolling task has received a growing attention from the multi-agent community due to its wide range of potential applications. Various algorithms based on reactive and cognitive architectures have been developed. However, the existing patrolling-specific approaches based on deep learning algorithms are still in preliminary stages. In this paper, we propose to integrate a recurrent neural network as part of a strategy for the multi-agent patrolling task. The recurrent neural networks and more specifically the LSTM architecture, as machines to learn temporal series, are well adapted to the multi-agent patrolling problem to the extent that the latter can be viewed as a decision problem over the time. In order to accomplish this study we proposed a formal model of an LSTM-based agent strategy called LSTM Path Maker . The LSTM network is trained over simulation traces of a fully-informed, coordinated and communicating strategy. Then each agent of the new strategy uses its LSTM network to select the next place to visit by feeding it with its current node. Finally this new LSTM-based strategy is evaluated in simulation and compared with two strategy, a cognitive and coordinated one, and a reactive and decentralised one. Preliminary results indicate that the proposed strategy is better than the decentralised one for the criteria of mean interval and quadratic mean interval, and but also close to HPCC for the former.