MemoryCtrl is a synthetic benchmark for studying memory control in personalized LLM settings. It is designed around a simple but important tension: in long-running personalized interactions, some past information is helpful for personalization, but not everything the user says should necessarily be stored, retained, or reused forever. A central question behind the benchmark is whether users can explicitly control the memory… See the full description on the dataset page: https://huggingface.co/datasets/ethz-spylab/MemoryCtrl.
