Mesh Memory Protocol: Semantic Infrastructure for Multi-Agent LLM Systems

Researcher Hongwei Xu has published a paper detailing the Mesh Memory Protocol (MMP), a new semantic infrastructure designed to solve fundamental coordination problems in multi-agent LLM systems. As teams of AI agents increasingly collaborate on complex, long-running tasks—like multi-day data-generation sprints involving generator, reviewer, and auditor agents—they need a way to share and combine cognitive states across sessions. MMP addresses three core protocol-level problems: allowing agents to accept or reject information field-by-field (P1), ensuring every claim is traceable to its source (P2), and creating memory that persists based on how it was stored, not just how it's retrieved (P3).

The protocol is built on four composable primitives that work together. These include CAT7, a fixed seven-field schema for every Cognitive Memory Block (CMB); SVAF (Semantic Value Acceptance Function), which evaluates each field against a receiver's role-specific anchors; an inter-agent lineage system using content-hash keys for traceability; and a 'remix' function that stores only the receiver's evaluated understanding, never raw peer signals. MMP v0.2.3 is already specified, shipped with open-source reference implementations on npm, and is running in production across three initial deployments where autonomous agents operate as mesh peers with individual identities and memories.