AIM Adds Index-Level Access Control to Agent Memory and Ships MUMBench for Multi-User Memory Operations

arXiv 2609.12320·medium signalXBlueskyLinkedInCopy link

Most agentic memory systems are scoped to one user, which blocks any sharing of genuinely public knowledge across users. AIM classifies each piece of information as private or public at write time and enforces index-level access control so private memories are retrievable only by their owner, and it comes with MUMBench, described as the first public dataset evaluating retrieval, creation, update, and deletion in a multi-user setting across four domains. Across three independent runs AIM hits 96.0% visibility-classification accuracy but only 58.8% strict operation accuracy and 70.5% state-aware operation accuracy, so the classification is the solved part and the operations are not.