A new study on foundation-model agents highlights the importance of deployment-time memorization, which affects personalization, extraction risk, and deletion fidelity. Key findings suggest that while summarization can enhance privacy, it may also lead to challenges in completely erasing sensitive information from memory systems.