AI memory helps personalize assistants, but users do not necessarily want every disclosed detail retained forever. A new arXiv paper takes up this problem by examining what it calls the "negative space" of AI memory: the areas where forgetting should be the default, even if that means losing potential personalization value.
The authors argue that in this negative space, protection should outrank future benefits from personalization. Rather than treating memory as a simple retention problem, they propose a design and taxonomy for governing when and how AI systems should forget.
Because the abstract is short, the paper's exact methods and examples are not detailed here. But the central claim is clear: AI assistants need an explicit governance layer for what they should not remember.