Governed Shared Memory Critical for Multi-Agent LLM Systems

Managing shared knowledge effectively in multi-agent large language model (LLM) environments presents significant challenges, which this paper formalizes as the "fleet-memory problem." The research identifies four critical failure modes: unauthorized information leakage, propagation of stale data, persistence of contradictory information, and collapse of provenance tracking. To address these issues, the authors define and implement explicit systems-level primitives within a production multi-tenant memory service called MemClaw. These primitives include scoped retrieval for access control, temporal supersession for managing data freshness, provenance tracking for accountability, and policy-governed memory propagation. The system was evaluated using ArgusFleet, a harness designed to test these governance dimensions. Key evaluation results showed MemClaw successfully reconstructed 100% of complex derivation chains with correct writer identity and achieved high intra-fleet visibility without cross-fleet leakage. The study also uncovered practical architectural issues, such as initial sub-tenant scope bypasses and pipeline ordering conflicts, which were subsequently remediated. The conclusion emphasizes that simple long-context retrieval is insufficient for production multi-agent memory, underscoring the need for explicit, governed shared memory architectures.