Deployment-Time Memorization in Foundation-Model Agents

The paper presents a study on deployment-time memorization in foundation-model agents, introducing a framework that evaluates memory design choices based on their impact on personalization utility and extraction risk. It introduces metrics such as Personalization Recall (PR), Adversarial Extraction Rate (AER), and the Forgetting Residue Score (FRS) to analyze memory configurations. Key findings include that summarization can significantly reduce data extraction risks while maintaining recall, but also highlight challenges in deletion fidelity, indicating that memory management must be treated as a critical aspect of agent design for ensuring user privacy and data security.