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 that span multiple sessions or involve specialist agents handing off findings can use MMP's four primitives as a concrete protocol blueprint for selective memory sharing, provenance tracking, and session-persistent cognitive state.
- 01Hongwei Xu introduces the Mesh Memory Protocol (MMP) as a 'semantic infrastructure' layer for multi-agent LLM systems.
- 02MMP targets 'cross-session agent-to-agent cognitive collaboration', explicitly distinguished from parallel agent execution.
- 03Three core problems are addressed: field-level selective acceptance (P1), full claim traceability (P2), and storage-based memory relevance (P3).
Hongwei Xu's paper identifies a missing layer in multi-agent LLM systems: a "semantic infrastructure" that sits above tool-access and task-delegation protocols and governs how agents share and evaluate each other's cognitive state across sessions. The motivating use cases include multi-day data-generation pipelines where generator, reviewer, and auditor agents coordinate on overlapping batches in real time, as well as specialist agents that must carry findings forward across session restarts and product decisions that compound over many review rounds.
The paper frames three protocol-level problems that must be solved together.
The paper frames three protocol-level problems that must be solved together. P1 requires that each agent accept or reject content field by field from peers, not as whole messages. P2 requires that every claim be traceable to its source so that returning claims are recognized as echoes of the receiver's own prior thinking. P3 requires that memory surviving session restarts be relevant because of how it was stored, not how it is retrieved. MMP addresses all three through four composable primitives: CAT7, a fixed seven-field schema for every Cognitive Memory Block (CMB); SVAF, which evaluates each field against the receiver's role-indexed anchors to realize P1; inter-agent lineage, carried as `parents` and `ancestors` of content-hash keys to realize P2; and remix, which stores only the receiver's own role-evaluated understanding of each accepted CMB — never the raw peer signal — to realize P3. The protocol is described as specified, shipped, and running in production across three reference deployments, where each session runs an autonomous agent as a mesh peer with its own identity and memory.
Key facts
- 01Hongwei Xu introduces the Mesh Memory Protocol (MMP) as a 'semantic infrastructure' layer for multi-agent LLM systems.
- 02MMP targets 'cross-session agent-to-agent cognitive collaboration', explicitly distinguished from parallel agent execution.
- 03Three core problems are addressed: field-level selective acceptance (P1), full claim traceability (P2), and storage-based memory relevance (P3).
- 04Four composable primitives form MMP: CAT7 (seven-field schema), SVAF (field-level evaluation), inter-agent lineage (content-hash-keyed provenance), and remix (role-evaluated storage).
- 05Every Cognitive Memory Block (CMB) uses the CAT7 fixed seven-field schema.
- 06The remix primitive stores only the receiver's role-evaluated understanding of accepted CMBs, never the raw peer signal.