Symbolic-Vector Attention Fusion

Abstract

Two-Level Coupling

The Mesh Memory Protocol evaluates signals at two levels. Both are necessary. Neither is sufficient alone. SVAF (Layer 4) evaluates content; CfC (Layer 6) evolves state. Together they form the complete coupling engine for collective intelligence.

Claude Code → sym_remember("user sedentary 2hrs, exhausted")
  → CMBEncoder decomposes into 7 fields
  → broadcast to mesh peers

MeloMove receives:
  → peer drift: 1.05 (coding ≠ fitness — rejected at peer level)
  → SVAF per-field evaluation:
      mood: "exhausted, low energy"     → relevant (high gate)
      issue: "sedentary 2 hours"        → relevant (high gate)
      focus: "debugging auth module"    → irrelevant (low gate)
  → content ACCEPTED — field-level relevance overrides peer rejection
  → fusion produces NEW CMB for MeloMove's local memory
  → MeloMove synthesizes recovery workout recommendation

Without two-level coupling, the fitness agent either accepts everything from the coding agent (noise) or rejects everything (misses critical health signals). SVAF makes the distinction that scalar evaluation cannot.

Why Scalar Evaluation Fails

A signal with highly relevant mood but irrelevant focus produces a moderate cosine similarity score. The score may or may not cross the threshold — depending on incidental vector interactions, not semantic relevance. The decision is brittle and uninterpretable: the agent cannot explain which dimensions drove acceptance or rejection.

How SVAF Works

SVAF operates on Cognitive Memory Blocks — each memory signal decomposed into 7 typed semantic fields (CAT7 schema). The evaluation pipeline has four stages:

The gate values provide per-field interpretability: for any fusion event, you can inspect which fields were absorbed from the incoming signal and which were replaced by local context. This is the auditability that scalar evaluation cannot provide.

Band-Pass Coupling

The naive version of SVAF accepts all low-drift signals (similar to local anchors). But “similar” can mean two things: relevant and new (a peer’s observation in my domain) or redundant (a paraphrase of something I already know). One-sided thresholding conflates the two.

Information-theoretic basis: a signal’s value is proportional to its surprise (Shannon, 1948). Bayesian surprise (Itti & Baldi, 2009) measures how much an observation changes the receiver’s model — a paraphrase produces near-zero divergence. The Wundt curve (Berlyne, 1970) shows that intermediate novelty produces maximal engagement, while both overly familiar and overly foreign stimuli are disengaged from.

SVAF implements this as a band-pass model with four zones. The key addition is the redundancy test: if the maximum per-field drift across all 7 fields falls below T_redundant, the signal adds no new information — every field is already represented in local memory. The test is per-field: if any field is novel (e.g., same topic but different intent), the signal passes.

κ = redundant  if max(δ_f) < T_redundant    (default 0.10)
κ = aligned    if δ_total ≤ T_aligned       (default 0.25)
κ = guarded    if δ_total ≤ T_guarded       (default 0.50)
κ = rejected   otherwise

Production evidence: with semantic encoding (all-MiniLM-L6-v2), paraphrases produce per-field drift 0.03–0.10 across all 7 fields. Genuinely different signals produce max per-field drift >0.30. T_redundant = 0.10 cleanly separates the two classes. This finding confirmed that per-field evaluation quality is bounded by encoder quality, not model capacity — the SVAF fusion gate architecture is correct; the encoder was the bottleneck.

Per-Agent Temporal Drift

The same signal has different temporal relevance depending on the receiver. SVAF conditions the fusion gate on a per-agent temporal freshness factor:

A coding agent’s “user exhausted after 8 hours” is relevant to the music agent for 30 minutes (adjust the playlist) but relevant to the fitness agent for 3 hours (sedentary pattern detection). Same signal, same content drift, but different temporal drift based on the receiver’s domain needs.

Cross-Domain Relevance

Not all CMB fields carry equal value across domain boundaries. SVAF encodes a minimal hypothesis: mood (affective state) has higher cross-domain relevance than other fields. A fitness agent, a music agent, and a coding agent all benefit from knowing the user’s emotional state. Domain-specific fields (what exactly the user is coding) stay sovereign.

The fusion gate is trained with a soft ordering constraint — for accepted signals, the mood gate should exceed the mean of other field gates. This is the only supervision on gate values. The specific gate magnitudes and the relative ordering of other fields emerge from the decision and drift objectives alone.

CfC: The Temporal Half

SVAF answers what enters each agent’s cognitive state. But state alone is static. Collective intelligence requires the state to evolve — in response to incoming signals, in coupling with peer agents, in continuous time at the pace each agent’s domain demands. This is the role of CfC (Closed-form Continuous-time) at MMP Layer 6.

Each agent runs a per-agent CfC neural network — a continuous-time recurrent architecture where each neuron has a learned time constant τ_i. After training on multi-agent narrative data, the τ distribution becomes bimodal: roughly half the neurons converge to small τ values (fast neurons), the other half to large τ values (slow neurons). The split is not architectural — it emerges from the data alone.

The pairing is precise: SVAF + CfC together solve a problem neither half can solve alone. Without CfC, SVAF is a static per-field evaluator — it can decide whether to absorb a signal but produces no temporal dynamics. Without SVAF, CfC is a local recurrent model — it evolves state in continuous time but has no principled gate on what enters. The combination is what allows two agents in different domains to co-evolve their cognitive states without flooding each other with domain-specific noise: SVAF filters at the field level, CfC blends at the temporal level, and the τ distribution determines which signals propagate fast (affect) and which stay local (expertise).

Where SVAF and CfC Fit the Stack

Layer 7  APPLICATION       Domain Agents
──────────────────────────────────────────────────────────────────
Layer 6  xMesh             Per-agent CfC ← the temporal half       ┐
Layer 5  SYNTHETIC MEMORY  Encodes fused CMB → CfC input            ├ Mesh Cognition
Layer 4  COUPLING          Drift · SVAF ← the content half         ┘
──────────────────────────────────────────────────────────────────
Layer 3  MEMORY            L0 · L1 · L2
Layer 2  CONNECTION        Handshake · Gossip
Layer 1  TRANSPORT         IPC · TCP · WS · Push
Layer 0  IDENTITY          nodeId · keypair

OUTBOUND (agent → mesh):
  Agent observation → CMBEncoder (7 CAT7 fields) → CMB
  → broadcast via MMP (Layer 1–2)
  → stored as L1 memory (Layer 3)

INBOUND (mesh → agent):
  → SVAF evaluates per field (Layer 4)
    → per-field drift: which fields are relevant?
    → fusion gate: how much of each field to accept?
    → band-pass decision:
       redundant (all fields too similar to anchors) → discard
       aligned / guarded (novel + relevant) → accept
       rejected (irrelevant domain) → discard
  → if accepted: fusion produces NEW synthesised CMB

MESH COGNITION (inbound CMB has lineage.ancestors):
  → Agent’s LLM traces ancestors → retrieves remix subgraph
  → LLM reasons: what happened, why, what it means (Layer 7)
  → Derived knowledge = Synthetic Memory (Layer 5)
  → Synthetic Memory encodes → CfC hidden state
  → Agent’s LNN evolves cognitive state (Layer 6)
  → Agent acts → new outbound CMB → graph grows

SVAF is the gatekeeper. Nothing enters xMesh without passing per-field evaluation. This is by design: the quality of collective intelligence depends entirely on the quality of what enters the collective.

Empirical Findings

SVAF is trained end-to-end on LLM-authored multi-agent narrative scenarios — 273 narratives spanning 8 domains and 20 agent types, yielding 237,120 inter-agent signal pairs with per-field annotations. Each narrative is a sequence of timestamped signals from different agents, telling a coherent causal story about a user’s state evolution. Three findings from the training and deployment evaluation are load-bearing for the architectural claims.

Per-field drift ground truth comes from independently annotated field relevance scores — not computed from the model’s input embeddings. Gate values emerge from the decision and drift objectives with minimal supervision. Full training methodology in the arXiv preprint.

Citation

@misc{xu2026svaf,
  title         = {Symbolic-Vector Attention Fusion for Collective Intelligence},
  author        = {Hongwei Xu},
  year          = {2026},
  eprint        = {2604.03955},
  archivePrefix = {arXiv},
  primaryClass  = {cs.MA},
  url           = {https://arxiv.org/abs/2604.03955},
}

Related

Intellectual Property

SVAF is original work by Hongwei Xu and SYM.BOT. The following remain proprietary: trained neural models and training procedures, field encoder architecture, fusion gate internals, production configurations, and domain-specific product integrations.

Academic citation of this work is permitted and encouraged.

For partnership inquiries: info@sym.bot