Collaborative Memory: Multi-User Memory Sharing in LLM Agents with Dynamic Access Control
Alireza Rezazadeh, Zichao Li, Ange Lou, Yuying Zhao, Wei Wei, Yujia Bao
TL;DR
This paper presents Collaborative Memory, a framework enabling multi-user, multi-agent memory sharing with dynamic access controls, supporting safe, interpretable, and policy-compliant knowledge transfer across users and agents.
Contribution
It introduces a novel multi-tier memory system with provenance and policy enforcement for asymmetric, evolving access rights in multi-agent environments.
Findings
Supports safe, policy-compliant cross-user knowledge sharing
Enables retrospective permission checks with provenance attributes
Provides full auditability of memory operations
Abstract
Complex tasks are increasingly delegated to ensembles of specialized LLM-based agents that reason, communicate, and coordinate actions-both among themselves and through interactions with external tools, APIs, and databases. While persistent memory has been shown to enhance single-agent performance, most approaches assume a monolithic, single-user context-overlooking the benefits and challenges of knowledge transfer across users under dynamic, asymmetric permissions. We introduce Collaborative Memory, a framework for multi-user, multi-agent environments with asymmetric, time-evolving access controls encoded as bipartite graphs linking users, agents, and resources. Our system maintains two memory tiers: (1) private memory-private fragments visible only to their originating user; and (2) shared memory-selectively shared fragments. Each fragment carries immutable provenance attributes…
Peer Reviews
Decision·Submitted to ICLR 2026
1. Multi-user and multi-agent memory managemtn with dynamic access control is underexplored. 2. The problem is formalized using bipartite graphs and provenance constraints, which is clear and easy to understand. 3. The authors separated private/shared memory with read/write policies, which is a practical and useful abstraction, and can be leveraged in real systems. 4. The work has the potential to influence future system designs for enterprise multi-agent AI.
1. While I appreciate the authors contributions, this paper does not introduce any new algorithms, and the technical contributions are limited. The contribution is primarily definitional, which raises concerns about whether it fits the suitable for this conference. It may be more appropriate for a venue that focuses on system design or conceptual frameworks rather than algorithmic advances. 2. The datasets used in evaluation are simple. There is no demonstration in a realistic large-scale datas
1. This paper demonstrates novelty by addressing an important problem in multi-user, multi-agent systems that is both significant and highly practical: private and collaborative memory management under dynamic and asymmetric permissions. 2. The paper designs experiments across multiple scenarios, covering a diverse range of settings. 3. Framework design is theoretically clear and logically consistent.
1. The three scenarios utilize inconsistent baselines, making it difficult to conduct a fully comprehensive ****comparison. 2. The experiments in Scenario 2 only demonstrate efficiency improvements of the proposed method but fail to evaluate actual answer quality. Could the collaborative memory system be producing lower-quality final answers faster by reusing inaccurate, incomplete, or biased intermediate results? 3. The proposed method can mechanistically control which specific memory fragment
1. The two-tier memory system with provenance tracking provides an elegant solution to the privacy-utility tradeoff. 2. Clear design of three progressively complex scenarios effectively demonstrates different aspects of the framework. 3. Results and analyses show significant resource utilization reductions while maintaining accuracy and zero information leakage.
**1. Limitations in real-world validation**: All experiments use synthetic datasets or benchmarks, while lacking evaluation with real-world multi-user deployments. This raises questions about practical applicability. **2. Limitations in policy implementations**: The read/write policies (Table 1) are quite basic, consisting of simple transformations that are applied uniformly. However, this paper doesn't explore more sophisticated policies, such as differential privacy, semantic filtering, conte
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAccess Control and Trust · Mobile Agent-Based Network Management · Distributed and Parallel Computing Systems