Governed Shared Memory for Multi-Agent LLM Systems
MemClaw addresses four failure modes in multi-agent LLM systems through governed shared memory, ensuring efficient knowledge management.
Key Findings
Methodology
The paper proposes a system architecture for governed shared memory, defining primitives for scoped retrieval, temporal supersession, provenance tracking, and policy-governed propagation. These are implemented in MemClaw and evaluated using ArgusFleet, revealing architectural issues in production environments.
Key Results
- In MemClaw, 100% of depth-four derivation chains were successfully reconstructed with correct writer identity at sub-second per-hop latency.
- Propagation tests showed high intra-fleet visibility with zero cross-fleet leakage, optimizing write-to-visible latency to a single search round-trip.
- Architectural issues such as asymmetric scope enforcement and pipeline ordering conflicts were discovered.
Significance
This study provides a comprehensive framework for memory governance in multi-agent LLM systems, addressing complexities unmanageable in single-agent environments. By introducing systems-level abstractions, it enhances memory management security and consistency, with significant academic and industrial implications.
Technical Contribution
The technical contribution lies in formalizing the shared memory problem as governed operational state, proposing new systems-level abstractions, and revealing design-missed issues through live production service evaluation.
Novelty
This is the first to formalize shared memory governance as operational state in multi-agent systems, empirically proving its effectiveness and addressing gaps in temporal contradiction resolution and supersession.
Limitations
- In GET-by-id requests, sub-tenant scope was initially unenforced, leading to scope leakage issues.
- A synchronous near-duplicate gate may reject contradictory writes before the asynchronous contradiction detector evaluates them.
Future Work
Future work can focus on optimizing the efficiency and scalability of memory governance, especially in large-scale multi-agent environments.
AI Executive Summary
The rise of multi-agent large language model (LLM) systems presents challenges in shared memory governance. Traditional single-agent memory systems cannot meet the complex demands of multi-agent environments, such as access control, temporal consistency, and provenance tracking.
This paper proposes a new system architecture called MemClaw, addressing these issues by defining primitives for scoped retrieval, temporal supersession, provenance tracking, and policy-governed propagation. Evaluated in a production environment, MemClaw demonstrates effectiveness in governing shared memory, particularly in preventing unauthorized leakage and ensuring efficient memory propagation.
However, the study also reveals architectural issues, such as asymmetric scope enforcement and pipeline ordering conflicts, indicating that design-missed problems require live evaluation to uncover. Future research directions include further optimizing the efficiency and scalability of memory governance to accommodate larger-scale multi-agent systems.
Deep Analysis
Background
With the widespread application of large language models, multi-agent systems are becoming mainstream. These systems require shared memory to coordinate behavior among different agents. However, traditional memory systems are designed primarily for single-agent environments and cannot meet the complex demands of multi-agent environments, such as access control, temporal consistency, and provenance tracking.
Core Problem
Shared memory governance in multi-agent systems faces four primary failure modes: unauthorized leakage, stale propagation, contradiction persistence, and provenance collapse. These issues not only affect system security and consistency but also lead to operational instability.
Innovation
The paper proposes a new system architecture by defining primitives for scoped retrieval, temporal supersession, provenance tracking, and policy-governed propagation to solve shared memory governance issues. These primitives are implemented in MemClaw and evaluated using ArgusFleet, demonstrating their effectiveness in a real production environment.
Methodology
- �� Scoped Retrieval: Ensures only authorized agents can access specific memory.
- �� Temporal Supersession: Resolves contradictory writes through temporal ordering and policy control.
- �� Provenance Tracking: Records writer identity and modification history for each memory object.
- �� Policy-Governed Propagation: Ensures safe memory propagation through policy filtering and temporal resolution.
Experiments
Experiments use ArgusFleet to evaluate MemClaw's performance, focusing on four governance dimensions: leakage, contradiction, provenance, and propagation. Each experiment is conducted in a real production environment to reveal design-missed issues.
Results
Experimental results show that MemClaw effectively prevents unauthorized leakage and ensures efficient memory propagation in all tests. Additionally, the experiments reveal architectural issues such as asymmetric scope enforcement and pipeline ordering conflicts.
Applications
The system can be used in multi-agent environments requiring efficient memory governance, such as enterprise collaboration systems, customer support platforms, and automated workflows.
Limitations & Outlook
Although MemClaw performs well in governing shared memory, there are limitations, such as initial unenforced sub-tenant scope in GET-by-id requests. Additionally, pipeline ordering conflicts may affect system performance.
Plain Language Accessible to non-experts
Imagine a library where multiple librarians need to manage and update book information simultaneously. Each librarian has their own permissions and responsibilities, ensuring the accuracy and consistency of information. MemClaw acts like an intelligent system that helps these librarians manage book information effectively, preventing information leakage and erroneous propagation.
ELI14 Explained like you're 14
Imagine you're playing a massive multiplayer online game with your friends, each having their own tasks and roles. MemClaw is like a super-smart game admin ensuring everyone gets the right info and doesn't fail missions due to wrong info. It's like an invisible helper making sure you all complete game tasks smoothly.
Glossary
Multi-Agent Systems
Systems composed of multiple independent agents that can collaborate to complete complex tasks.
In this paper, multi-agent systems require shared memory to coordinate behavior.
Shared Memory
Memory shared among multiple agents for storing and exchanging information.
Shared memory is key to coordinating agent behavior in multi-agent systems.
Governance
The process of controlling system behavior through policies and rules to ensure security and consistency.
The paper proposes a system architecture for governing shared memory.
Provenance Tracking
The process of recording the source and modification history of information to ensure traceability.
Provenance tracking is crucial for ensuring the auditability of memory systems.
Temporal Supersession
A method of resolving information contradictions through temporal ordering to ensure consistency.
Temporal supersession addresses contradictory writes in multi-agent systems.
Open Questions Unanswered questions from this research
- 1 How to effectively govern shared memory in larger-scale multi-agent environments remains to be further studied.
- 2 Existing methods are inefficient in handling complex temporal contradictions and need optimization.
Applications
Immediate Applications
Enterprise Collaboration Systems
Helps enterprises share and manage information across departments, ensuring information security and consistency.
Long-term Vision
Automated Workflows
In future automated systems, MemClaw can serve as a core component, helping manage complex tasks and information flows.
Abstract
Multi-agent LLM environments require robust mechanisms for shared knowledge management. This paper formalizes the fleet-memory problem and identifies four foundational failure modes: unauthorized leakage, stale propagation, contradiction persistence, and provenance collapse. To address these, we define explicit systems-level primitives: scoped retrieval, temporal supersession, provenance tracking, and policy-governed memory propagation. These primitives are implemented in MemClaw, a production multi-tenant memory service, and evaluated via ArgusFleet, a reproducible harness testing four governance dimensions. Rather than a baseline comparison, this study measures a live production service, emphasizing real-world architectural insights and negative results. Key Evaluation Results Provenance: Successfully reconstructed 100% of depth-four derivation chains with correct writer identity at sub-second per-hop latency. Propagation: Demonstrated high intra-fleet visibility with zero cross-fleet leakage. Under strong write mode, write-to-visible latency was optimized to a single search round-trip. Production Architectural Issues Discovered Asymmetric Scope Enforcement: Tenant isolation held, but sub-tenant scope was initially bypassed on direct GET-by-id requests for agent-scoped credentials (disclosed and remediated during the study). Pipeline Ordering Conflict: While contradiction supersession works for admitted writes, a synchronous near-duplicate gate can prematurely reject contradictory writes before the asynchronous contradiction detector can evaluate them. Conclusion: Long-context retrieval alone is insufficient for production multi-agent memory. Governed shared memory demands explicit systems-level abstractions, and live evaluation is vital to expose enforcement and pipeline-ordering failures missed by design-only treatments.