Continual learning protocols are attracting increasing attention from the\nmedical imaging community. In continual environments, datasets acquired under\ndifferent conditions arrive sequentially; and each is only available for a\nlimited period of time. Given the inherent privacy risks associated with\nmedical data, this setup reflects the reality of deployment for deep learning\ndiagnostic radiology systems. Many techniques exist to learn continuously for\nimage classification, and several have been adapted to semantic segmentation.\nYet most struggle to accumulate knowledge in a meaningful manner. Instead, they\nfocus on preventing the problem of catastrophic forgetting, even when this\nreduces model plasticity and thereon burdens the training process. This puts\ninto question whether the additional overhead of knowledge preservation is\nworth it - particularly for medical image segmentation, where computation\nrequirements are already high - or if maintaining separate models would be a\nbetter solution. We propose UNEG, a simple and widely applicable multi-model\nbenchmark that maintains separate segmentation and autoencoder networks for\neach training stage. The autoencoder is built from the same architecture as the\nsegmentation network, which in our case is a full-resolution nnU-Net, to bypass\nany additional design decisions. During inference, the reconstruction error is\nused to select the most appropriate segmenter for each test image. Open this\nconcept, we develop a fair evaluation scheme for different continual learning\nsettings that moves beyond the prevention of catastrophic forgetting. Our\nresults across three regions of interest (prostate, hippocampus, and right\nventricle) show that UNEG outperforms several continual learning methods,\nreinforcing the need for strong baselines in continual learning research.\n