In this paper, we propose a semi-supervised setting for semantic segmentation of a whole volume from only a tiny portion of one slice annotated using a memory-aware network pre-trained on video object segmentation without additional fine-tuning. The network is modified to transfer annotations of one partially annotated slice to the whole slice, then to the whole volume. This method discards the need for training the model. Applied to Electron Tomography, where manual annotations are time-consuming, it achieves good segmentation results considering a labeled area of only a few percent of a single slice. The source code is available at https://github.com/licyril1403/hijacked-STM.