The abstract of "Let Them Sleep: Recursive Self-Pruning Memory for Long-Horizon LLM Agents" (arXiv:2610.25561) identifies a specific failure mode for agents that run for a long time. It argues that the memory problem such agents face is not the one most systems optimize for. Storing every trace and retrieving nearest neighbors works early, but quietly stops working later as the store fills up.

The paper's title suggests the proposed remedy: recursive self-pruning memory. Rather than simply accumulating more data, the agent would prune its own memory in a structured way, allowing it to continue operating over long horizons. The abstract does not provide further details in the portion available.