Long-term memory is now a standard component of large language model (LLM) agents, and most research has concentrated on making retrieval faster and more accurate. But a new preprint, MemCut, argues that this focus misses half the problem: a dependable memory system must also be able to forget.
The paper, posted on arXiv, does not provide full details in its abstract, but its central claim is clear. It positions forgetting as a necessary complement to retrieval, suggesting that without a way to remove or suppress outdated or harmful information, an agent's memory becomes less trustworthy over time.
MemCut appears to be a proposed mechanism for introducing this capability, though the abstract leaves the exact method unspecified. The significance lies in shifting attention from purely additive memory to a more balanced view, where deletion is as important as storage. As the authors put it, every secret leaves a trail in long-term memory—and sometimes that trail needs to be cut.