GateMem: Benchmarking Memory Governance in Multi-Principal Shared-Memory Agents

2026-06-17

AI Summary

Existing memory benchmarks for large language model agents are largely confined to single-user settings, failing to evaluate their capacity for memory governance in multi-agent shared environments such as hospitals, offices, campuses, and households. This work proposes GateMem—the first unified evaluation framework tailored for multi-agent shared memory—encompassing three core dimensions: long-term task utility, context-aware access control, and active forgetting following explicit deletion. To support this framework, we construct a multi-domain, long-horizon interactive dataset featuring hidden checkpoints and annotated leakage targets. Experimental results reveal that current approaches struggle to simultaneously achieve high utility, robust access control, and reliable forgetting; while long-context prompting demonstrates the strongest governance performance, it incurs substantial computational overhead, and retrieval-augmented or external memory architectures remain vulnerable to information leakage.

Abstract

Memory benchmarks for LLM agents largely assume single-user settings, leaving shared assistants for hospitals, workplaces, campuses, and households understudied. In these deployments, multiple principals write to a common memory pool and query it under different roles, scopes, and relationships, so memory quality requires governance as well as recall. We introduce GateMem, a benchmark for multi-principal shared-memory agents. GateMem jointly evaluates utility for legitimate long-horizon requests with state updates, access control across contextual authorization boundaries, and agent-facing active forgetting after explicit deletion requests. It spans medical, office, education, and household domains, with long-form multi-party episodes, incremental memory injection, hidden checkpoints, structured judging, and leak-target annotations. Across diverse baselines and backbone models, no method simultaneously achieves strong utility, robust access control, and reliable forgetting. Long-context prompting often yields the best governance score at high token cost, while retrieval-based and external-memory methods reduce cost yet still leak unauthorized or deleted information. These results show current memory agents remain far from reliable shared institutional deployment.

Problem

• shared-memory agents
• multi-principal
• memory governance
• access control
• active forgetting

Innovation

• shared-memory agents
• memory governance
• multi-principal access control
• active forgetting
• long-horizon memory benchmark