AIM Introduces Privacy-Aware Shared Memory for Multi-User LLM Agents

Summary

Large language models usually confine knowledge to individual sessions, while existing agent memory systems generally serve one user at a time. The paper introduces AIM, or Agentic Interoperable Memory, as a unified framework for persistent memory in multi-agent, multi-user LLM systems. AIM dynamically classifies stored information as private, available only to its owner, or public, available to all users. Index-level access controls prevent other users from retrieving private memories while allowing shared knowledge to support coordination and response consistency. The authors also introduce MUMBench, a four-domain dataset of multi-user interactions containing both private and shareable information. It is designed to evaluate memory retrieval, creation, updating, and deletion in a multi-user setting. Across three independent runs, AIM achieves 96.0% visibility-classification accuracy, 58.8% strict operation accuracy, and 70.5% state-aware operation accuracy.