Mesh Memory Protocol (MMP)

• Mesh Memory Protocol (MMP) is a semantic infrastructure enabling real-time, field-granular cognitive exchange among autonomous agents.
• Its layered design integrates Cognitive Memory Blocks with Symbolic-Vector Attention Fusion and strict privacy guarantees to support resilient multi-agent collaboration.
• MMP employs lineage tracing and write-time filtering to prevent echo loops and ensure robust, persistent cognitive state continuity.

The Mesh Memory Protocol (MMP) is a semantic-infrastructure protocol designed for real-time, field-granular exchange and integration of cognitive state among autonomous agents, such as LLMs or on-device inference systems. By introducing a structured schema for high-level memory (Cognitive Memory Blocks, CMBs), per-field admission via Symbolic-Vector Attention Fusion (SVAF), robust lineage tracing, and a write-time filtering invariant (remix), MMP enables peer-to-peer, privacy-preserving, and resilient cross-session agent collaboration and cognitive coupling. Its layered architecture, strict privacy guarantees, and deployment in both multi-agent LLM environments and consumer on-device applications establish MMP as a foundational substrate for distributed multi-agent intelligence (Xu, 21 Apr 2026, Xu, 12 Apr 2026).

1. Design Rationale and Protocol Objectives

MMP addresses three tightly-coupled requirements for multi-agent collaboration: (P1) Per-field semantic admission: Agents must be able to evaluate and accept or reject individual semantic fields in peer messages, instead of coarse message- or session-level acceptance. (P2) Signal-level provenance and echo detection: Full traceability of each claim—agents can detect and suppress “echoes” of their own previous statements or decisions. (P3) Persistent, filter-relevant memory: Memory surviving agent restarts is constructed at write-time via role-based filtering, guaranteeing that only semantically meaningful, agent-specific knowledge is retained and restored.

MMP is explicitly privacy-oriented: in peer-to-peer mesh deployments, all inference remains local to each autonomous peer. No hidden states, model weights, or gradients leave the device or agent. Instead, only structured CMBs representing semantically audited, role-indexed high-level fields are exchanged. Echo loops are systematically prevented through lineage-based rejection, while bounded-drift and layer separation protect against semantic incoherence and private-state leakage (Xu, 12 Apr 2026).

2. Layered Architecture and Core Primitives

MMP is instantiated as a multi-layer protocol stack, with semantic infrastructure comprising Layers 3–4 (broadcast and field evaluation), and critical cognitive dynamics at higher layers. The core protocol primitives are:

CAT7 Schema: Every CMB is a fixed seven-field record (focus, issue, intent, motivation, commitment, perspective, mood), each field containing both symbolic (text) and embedding (unit-normalized $\mathbb{R}^d$ vector) content. Mood fields may additionally carry affective coordinates (valence, arousal). Field uniformity enables universal coupling and eliminates per-domain schema negotiation (Xu, 21 Apr 2026).

SVAF (Symbolic-Vector Attention Fusion): At Layer 4, per-field drift $\delta_f = 1 - \cos(\hat v_f, v_{\text{new}}^f)$ is computed between incoming field vectors and local anchor memories. Admission is determined by band-pass thresholds:

• Redundant ($\max_f \delta_f < T_{\text{red}}$): drop as already seen,
• Aligned ($\delta_{\text{total}} \leq T_{\text{aln}}$): full fusion,
• Guarded ($\delta_{\text{total}} \leq T_{\text{grd}}$): attenuated fusion,
• Rejected ($\delta_{\text{total}} > T_{\text{grd}}$): block (Xu, 21 Apr 2026, Xu, 12 Apr 2026). SVAF operates independently per agent, with role-indexed field weights and memory freshness decay.

Lineage and Provenance: Every CMB carries content-hash keys, as well as explicit parents and recursively computed ancestors sets. Incoming CMBs whose ancestry intersects the receiver’s key-set are recognized as echoes and dropped in $O(1)$ time (Xu, 21 Apr 2026). This enables causality tracing and echo loop suppression.

Remix/Write-Time Filtering: Aligned or guarded CMBs are not stored verbatim. Instead, the receiving agent produces a new CMB with its own key, filtered header, preserved lineage, and optional body. This invariant ensures memory relevance and agent-specificity upon session restart; raw peer signals or uninterpreted transcripts are never retained (Xu, 21 Apr 2026).

3. Continuous-Time Cognitive Dynamics

MMP accommodates both episodic and real-time distributed cognition through the integration of closed-form continuous-time (CfC) neural networks. In production deployments such as MeloTune, each peer maintains two independent CfC networks:

• Listener-level CfC: Private, per-device ODE-based model ingesting track context signals (valence, arousal, etc.), capturing individualized affective trajectories.
• Mesh-runtime CfC (Layer 6): Consumes fused, SVAF-admitted CMB fields and updates a shared memory state using the update:

$$h(t+\Delta t) = h(t) \odot e^{-\Delta t/\tau} + (1 - e^{-\Delta t/\tau}) \odot f_{\theta}([x(t), h(t)])$$

Different time constants $\tau_{\text{mesh}}$ parameterize fast vs. slow neurons, mediating rapid mesh-wide synchronization and private, long-term expertise.

The scalar “coherence” value $\rho(t) \in [0, 1]$ quantifies the convergence of mesh mood around shared semantic/affective context and can directly influence downstream logic (e.g., music curation) (Xu, 12 Apr 2026).

4. Message Format, Protocol Operation, and Privacy Guarantees

MMP communication is fully peer-to-peer. Peers discover each other via dynamic protocols (Bonjour), then exchange binary CMB packets over UDP multicast. CMB packets include the CAT7 fields, lineage information, and (optionally) a body schema. Redundancy and freshness (∼30 min window) are enforced in protocol layers, and echo-prone CMBs are programmatically suppressed via lineage keys.

All inference—both CfC updates and SVAF gating—executes locally (e.g., via CoreML) with sub-millisecond latency. Hidden states, model weights, and training signals (e.g., Personal Arousal Function [PAF] adaptation) remain device-resident, processed via local accumulators (EMA tables), and are never exchanged or synchronized, fulfilling strict R4 protocol guarantee (“no hidden states on the wire”). Organic-mood constraints isolate mesh-induced state updates for a fixed interval to prevent recursive resubmission and induce robustness to adversarial or spurious coupling (Xu, 12 Apr 2026).

5. Semantic Admission, Echo Loop Prevention, and Application in Collaboration

The SVAF-based field evaluation and write-time remix underpin three properties specific to cross-agent cognition in long-running and restarted collaborations:

• Per-field semantic agreement: Agents can partially accept peer CMBs, e.g., importing mood fields while guarding or rejecting commitment fields if those conflict with role-indexed anchors, eliminating all-or-nothing gating and schema drift (Xu, 21 Apr 2026).
• Echo detection: Lineage ancestry walk identifies recapitulation of one’s own prior claims for immediate suppression—even in multi-round deliberations, preventing recursive consensus loops.
• Session persistence: On agent restart, work state is rebuilt from the local mesh memory which only contains role-filtered, remixed CMBs; no history replay or raw transcript exchange is required.

In production settings (e.g., MeloTune, Anthropic Claude Code with sym-mesh-channel plugin), these properties facilitate seamless, low-latency agent orchestration with strong domain-invariant schema (CAT7) and robust semantic filtering.

6. Reference Deployments, Evaluations, and Observed Performance

MMP is operational across diverse agent architectures:

Deployment	Substrate/System	Observed Results
MeloTune (iOS)	CoreML, SYM Swift peer	Trajectory MAE 0.414; $\delta_f = 1 - \cos(\hat v_f, v_{\text{new}}^f)$00.6 ms/step; 95k params; mood convergence $\delta_f = 1 - \cos(\hat v_f, v_{\text{new}}^f)$1 in 5–10 s (Xu, 12 Apr 2026)
Anthropic Claude plugin	Claude Channels + MMP plugin (sym-mesh)	78.7% three-class SVAF accuracy; echo loops prevented, restart latency $\delta_f = 1 - \cos(\hat v_f, v_{\text{new}}^f)$2 s (Xu, 21 Apr 2026)
Multi-agent mesh sprint	Multi-role mesh, 14-wave narrative	78.2% retention (gate); 98.5% compliance after fix

In all reference cases, MMP enables robust peer-to-peer cognitive-state exchange, low-bandwidth field fusion (hundreds of CMBs/minute, $\delta_f = 1 - \cos(\hat v_f, v_{\text{new}}^f)$3 ms latency), and seamless continuity across agent restarts. Qualitatively, shared mood coupling and intent propagation occurs on sub-10 s time scales. No privacy compromise or hidden-state leakage has been observed (Xu, 12 Apr 2026, Xu, 21 Apr 2026).

7. Significance and Outlook

MMP establishes, for the first time, a fully specified, implemented, and production-deployed protocol for semantic infrastructure among autonomous agents. Its impact is twofold:

• In multi-agent LLM systems, it provides principled mechanisms for modular collaboration and durable provenance, essential for verification, auditability, and reliable meta-coordination over extended time horizons.
• For edge/consumer inference, it demonstrates local, privacy-respecting mood and experience coupling (e.g., cross-device music curation) with near-real-time dynamical adaptation and zero cross-agent model sharing.

A plausible implication is increased adoption in domains requiring multi-agent continuity, auditability, and privacy guarantees. The protocol’s extensibility (e.g., to arbitrary body schemas, dynamic roles, or novel cognitive domains), while preserving protocol invariants, remains a key open direction for research and deployment (Xu, 21 Apr 2026, Xu, 12 Apr 2026).