Paper Overview
Field: Machine Learning Authors: Zachary Johnson, Nigel Boachie Kumankumah, Somya Chatterjee, Tejas Sathyamurthi, Min Chen, Xinyi Alice Li, Xiao Wang, Emily Morgan Gelchie, Jessica Lin, Sadid A. Hasan, Sulaiman Vesal Published: 2026-09-15 arXiv: 2609.12320
Summary
Traditional large language models (LLMs) are scoped to individual user sessions, limiting their knowledge to a single conversation and preventing them from learning user preferences that evolve over time. Existing agentic memory systems address this limitation but generally operate at the individual-user level, restricting the public knowledge that could be shared across users to improve downstream responses.
The authors introduce AIM (Agentic Interoperable Memory), a unified, privacy-aware memory framework that enables multi-agent, multi-user LLM systems to persistently manage private and shared memory. AIM dynamically classifies information as:
- Private: scoped to one user and inaccessible to others
- Public: accessible to all users
- 96.0% visibility classification accuracy
- 58.8% strict operation accuracy
- 70.5% state-aware operation accuracy
AIM enforces index-level access controls so that private memories can only be retrieved by their owner—protecting sensitive data while allowing beneficial shared knowledge to improve coordination and consistency.
MUMBench Benchmark
The paper also proposes MUMBench (Multi-User Memory Benchmark), a multi-user interaction dataset containing private and shareable information across four domains. To the authors' knowledge, MUMBench is the first public dataset designed to evaluate multiple memory operations—retrieval, creation, update, and deletion—in multi-user environments.
Results
Across three independent runs on MUMBench, AIM achieved: