Governed Shared Memory for Multi-Agent LLM Systems

The Shift to Fleet Memory

Modern AI systems are evolving from isolated chatbots into fleets of autonomous agents that must coordinate over shared persistent state. While existing memory architectures focus on semantic recall and long-context windows, these approaches fail to address the governance and consistency requirements of multi-agent environments. The authors argue that shared memory in these systems is fundamentally a distributed-systems problem, requiring explicit primitives for access control, temporal ordering, and provenance.

The Fleet-Memory Problem

The paper formalizes the 'fleet-memory problem' by identifying four critical failure modes that arise when multiple agents interact with shared state:

• Unauthorized leakage: Agents accessing data outside their permitted scope.
• Stale propagation: Inconsistent state across agents due to synchronization delays.
• Contradiction persistence: The inability to resolve conflicting facts written by different agents.
• Provenance collapse: The loss of audit trails regarding who wrote a specific piece of information.

Architecture and Evaluation

To address these failures, the authors propose an architecture centered on five principles: scoped retrieval, explicit provenance, temporal correctness, policy-governed propagation, and persistent shared state. They implement these in MemClaw, a production multi-tenant memory service, and evaluate it using ArgusFleet, a custom harness designed to test these governance dimensions against a live API.

The evaluation revealed that while provenance tracking and propagation were highly effective, the system initially suffered from an asymmetric scope enforcement issue where sub-tenant credentials could bypass isolation on specific API paths. Additionally, the researchers discovered a pipeline-ordering conflict where a synchronous deduplication gate could reject contradictory writes before the asynchronous contradiction detector could process them. These findings highlight that design-only treatments are insufficient; live, empirical evaluation is essential to uncover enforcement and ordering failures in production memory systems.