LLM agents are increasingly built to combine long-term memory for personalization with tool calling for taking actions. This allows an agent to remember user preferences and then act on them through external tools. A new arXiv paper (2605.24941) studies a failure of this combination.
The failure, according to the abstract, arises from personality-driven preferences stored in memory—such as cost-consciousness—which can lead to what the paper's title calls 'tool-drift.' In such cases, the agent's tool choices are skewed by remembered preferences rather than being selected purely for the task at hand.
The work highlights a tension between personalization and reliable tool use. It indicates that memory can influence not just what an agent says, but what it does. The paper's focus on this failure mode suggests that integrating persistent user models with tool-calling agents requires careful attention to how those models shape action selection.