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

Abstract: Teams of LLM agents increasingly collaborate on tasks spanning days or weeks: multi-day data-generation sprints where generator, reviewer, and auditor agents coordinate in real time on overlapping batches; specialists carrying findings forward across session restarts; product decisions compounding over many review rounds. This requires agents to share, evaluate, and combine each other's cognitive state in real time across sessions. We call this cross-session agent-to-agent cognitive collaboration, distinct from parallel agent execution. To enable it, three problems must be solved together. (P1) Each agent decides field by field what to accept from peers, not accept or reject whole messages. (P2) Every claim is traceable to source, so returning claims are recognised as echoes of the receiver's own prior thinking. (P3) Memory that survives session restarts is relevant because of how it was stored, not how it is retrieved. These are protocol-level properties at the semantic layer of agent communication, distinct from tool-access and task-delegation protocols at lower layers. We call this missing protocol layer "semantic infrastructure," and the Mesh Memory Protocol (MMP) specifies it. Four composable primitives work together: CAT7, a fixed seven-field schema for every Cognitive Memory Block (CMB); SVAF, which evaluates each field against the receiver's role-indexed anchors and realises P1; inter-agent lineage, carried as parents and ancestors of content-hash keys and realising P2; and remix, which stores only the receiver's own role-evaluated understanding of each accepted CMB, never the raw peer signal, realising P3. MMP is 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, collaborating with other agents across the network for collective intelligence.

• The paper introduces a novel semantic infrastructure using a universal CAT7 schema, role-indexed SVAF, lineage DAG, and remix memory to enable persistent multi-agent collaboration.
• It achieves robust performance with 78.2% narrative retention and 98.5% convergence pass rates in real-world, three-agent deployments.
• The protocol overcomes limitations of previous frameworks by ensuring field-specific semantic admission, provenance tracking, and contextually relevant memory filtering.

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

Motivation and Protocol-Level Requirements

As the deployment of multi-agent LLM systems on production workloads evolves, coordination across persistent, multi-session agent teams increasingly demands robust protocol-level semantic infrastructure. Such infrastructure is necessary to enable cross-session agent-to-agent cognitive collaboration, allowing agents to share, evaluate, and combine cognitive state in real time and across session boundaries. Prevailing frameworks and protocols have overlooked key requirements: per-field semantic admission (P1), signal-level lineage for claim provenance and echo detection (P2), and write-time filtering for relevance at admittance rather than retrieval (P3). The Mesh Memory Protocol (MMP) addresses these protocol-layer requirements by introducing four composable primitives: CAT7 field schema, Symbolic-Vector Attention Fusion (SVAF) evaluation, inter-agent lineage DAG, and write-time-filtered remix memory.

Protocol Architecture and Semantic Primitives

MMP is formally specified as an eight-layer architectural stack, separating protocol infrastructure (identity, transport, connection, memory) and mesh cognition (coupling, synthetic memory, per-agent Liquid Neural Networks, and application-specific reasoning). The semantic-infrastructure primitives operate primarily at Layers 3 and 4, bridging memory persistence and role-indexed field evaluation.

Figure 1: MMP's 8-layer architecture, with CAT7, SVAF, lineage, and remix primitives located in Layers 3 and 4 to enable semantic cognitive collaboration across autonomous agent sessions.

CAT7: Universal Seven-Field Cognitive Memory Schema

Each cognitive exchange in MMP is represented as a Cognitive Memory Block (CMB) structured by the fixed CAT7 schema: focus, issue, intent, motivation, commitment, perspective, mood. This schema enables domain-agnostic, typed semantic evaluations whereby field-specific content (symbolic text and embedding vectors) can be universally parsed and assessed. The header-body separation ensures schema invariance for semantic evaluation via SVAF, while supporting agent-authored task-specific body payloads for richer structured exchanges.

SVAF: Role-Indexed Per-Field Evaluation and Admission

Symbolic-Vector Attention Fusion (SVAF) enables fine-grained per-field evaluation, computing drift between incoming field vectors and receiver anchors. Field weights indexed to recipient agent roles produce variable aggregate drifts and contribute to band-pass classification into redundant, aligned, guarded, or rejected states. This approach admits high-novelty signal based on semantic alignment through role-specific field importance schedules, satisfying the requirement for receiver-autonomous field-granular admission and supporting advanced coordination mechanisms.

Lineage: DAG Encoding for Provenance and Echo Detection

CMBs are fully threaded with parent and ancestor references, establishing signal-level provenance and enabling $O(1)$ echo detection via content-hash key intersection. This directly addresses the degeneration of thought and recursive amplification failure modes documented in prior multi-agent debate frameworks, providing robust claim-grounding and retention primitives that span agent and session boundaries.

Remix: Write-Time Filtering and Persistent Evaluated Memory

MMP enforces a write-time filtering invariant whereby memory stores contain only receiver-remixed CMBs reflecting their domain-filtered understanding, never raw peer signals. This mechanism ensures contextually relevant recall and fast resumption post-session restart, inverting standard RAG-style read-time filtering approaches. Role-specialised context continuity is achieved by construction, not via retrieval heuristics, directly solving the persistent memory relevance problem.

Figure 2: MMP mesh topology showing cross-platform agent sessions coordinating over agent-to-agent channels, with CMBs routed, evaluated, and remixed via protocol-layer primitives.

Implementation and Operational Observations

The protocol is fully implemented and deployed across reference applications, including a Claude-native channel plugin and a cross-domain consumer mesh agent. Operational deployments on a three-agent mesh (execution, quality review, compliance) exhibit key protocol behaviors: methodology convergence via lineage-traced re-reading with zero human arbitration, cross-session context resumption without replay through persistent remix memory, and agent-authored structured body exchanges leveraging CAT7's header-body affordance.

Empirical results from a 14-wave corpus-generation sprint report a retained narrative rate of 78.2% and post-convergence pass rates averaging 98.5% across waves, with protocol primitives delivering halt-diagnose-fix convergence and robust memory continuity across session restarts.

Comparative Analysis and Theoretical Implications

Audit of prevailing memory substrates (MemGPT, Mem0, A-MEM, Reflexion, CoALA, Voyager, Collaborative Memory) shows that none jointly satisfies P1 (per-field admission), P2 (signal lineage), P3 (write-time filtering), and multi-agent first-class support. Collaborative Memory is the closest conceptual neighbor but diverges on access-control vs role-indexed semantic evaluation.

MMP introduces a necessary protocol layer for structured collective intelligence, with vertical composition (memory to cognition within agents) and horizontal composition (evaluated cognitive state across agents and sessions) realized via the four primitives. The protocol is agnostic to underlying model provider and supports agent heterogeneity, though cross-provider mesh empirical evaluations remain open work.

Limitations and Extensions

Current deployments are bounded to $N=3$ agents, same-provider, and single-team geography. Formal scalability and adversarial resilience under larger meshes, cryptographic per-CMB signing, and learned or topic-dependent $\alpha_f$ weight schedules are open areas for analytical extension. Integration of experiential-tier playbook induction and harmonization with local memory substrates are downstream applications enabled by MMP's semantic infrastructure but not directly addressed in the protocol itself.

Conclusion

Mesh Memory Protocol introduces semantic infrastructure essential for multi-agent LLM systems, formalizing universally typed field schema, role-indexed per-field evaluation, signal-level lineage for provenance, and write-time-filtered remix memory. These composable primitives provide a robust substrate for persistent, grounded, and autonomous cognitive collaboration across agent teams and session boundaries, meeting requirements unaddressed by prior orchestration and memory frameworks. Formal analysis of minimality, uniqueness, and scalability, as well as broader deployment in heterogeneous agent meshes, are promising directions for extension.