Mesh Memory Protocol defines semantic layer for multi-agent LLM collaboration
Hongwei Xu introduces the Mesh Memory Protocol (MMP), a semantic infrastructure specification enabling LLM agents to share, evaluate, and persist cognitive state across sessions and agent boundaries.
Score breakdown
Teams building multi-agent systems can reference MMP as a concrete protocol specification for persistent, traceable, and selectively integrated shared memory — addressing a gap that tool-access and task-delegation frameworks do not cover.
- 01Hongwei Xu introduces the Mesh Memory Protocol (MMP), a semantic infrastructure layer for multi-agent LLM systems.
- 02MMP targets 'cross-session agent-to-agent cognitive collaboration', distinguished from parallel agent execution and task-delegation protocols.
- 03Problem P1: agents must accept or reject memory field-by-field, not whole messages.
Hongwei Xu's paper identifies a missing layer in multi-agent LLM systems: a semantic infrastructure that lets agents share and integrate each other's cognitive state across sessions, not just exchange messages or delegate tasks. The paper frames this as "cross-session agent-to-agent cognitive collaboration," motivated by real workflows such as multi-day data-generation sprints where generator, reviewer, and auditor agents coordinate on overlapping batches, or product decisions that compound across many review rounds.
MMP is built around three protocol-level properties and four composable primitives.
MMP is built around three protocol-level properties and four composable primitives. The three properties are: (P1) field-by-field selective acceptance of peer memory rather than whole-message accept/reject; (P2) full inter-agent lineage so that every claim is traceable to its source, preventing agents from treating echoes of their own prior thinking as new external information; and (P3) session-persistent memory whose relevance is determined by how it was stored, not how it is retrieved. The four primitives that realize these properties are `CAT7` (a fixed seven-field schema for every Cognitive Memory Block, or CMB), `SVAF` (which evaluates each field against the receiver's role-indexed anchors, realizing P1), inter-agent lineage carried as `parents` and `ancestors` of content-hash keys (realizing P2), and `remix` (which stores only the receiver's role-evaluated understanding of an accepted CMB, never the raw peer signal, realizing P3). The paper reports that MMP is specified, shipped, and running in production across three reference deployments.
Key facts
- 01Hongwei Xu introduces the Mesh Memory Protocol (MMP), a semantic infrastructure layer for multi-agent LLM systems.
- 02MMP targets 'cross-session agent-to-agent cognitive collaboration', distinguished from parallel agent execution and task-delegation protocols.
- 03Problem P1: agents must accept or reject memory field-by-field, not whole messages.
- 04Problem P2: every claim must be traceable to its source via inter-agent lineage using content-hash keys with `parents` and `ancestors` fields.
- 05Problem P3: session-persistent memory must be relevant based on how it was stored, not how it is retrieved.
- 06MMP comprises four composable primitives: `CAT7` (seven-field schema), `SVAF` (role-indexed field evaluation), inter-agent lineage, and `remix` (stores only role-evaluated understanding, never raw peer signal).