During the $\mu-3e$ experiment we faced the challenge of reconstructing the paths of certain low momentum particles that curled back into the detector and cause additional hits. To face this, a recurrent neural network was used which found the right track for $87 \%$ of these particles.