Symbolic-Vector Attention Fusion for Collective Intelligence

Hongwei Xu

cs.MAApr 5, 2026

Symbolic-Vector Attention Fusion for Collective Intelligence

Authors: Hongwei Xu

Abstract

When autonomous agents observe different domains of a shared environment, each signal they exchange mixes relevant and irrelevant dimensions. No existing mechanism lets the receiver evaluate which dimensions to absorb. We introduce Symbolic-Vector Attention Fusion (SVAF), the content-evaluation half of a two-level coupling engine for collective intelligence. SVAF decomposes each inter-agent signal into 7 typed semantic fields, evaluates each through a learned fusion gate, and produces a remix -- new knowledge from the intersection of two domains. A band-pass model yields four outcomes (redundant, aligned, guarded, rejected), solving both selectivity and redundancy. The fusion gate independently discovers a cross-domain relevance hierarchy: mood emerges as the highest-weight field by epoch 1, before accuracy plateaus -- consistent with independent mechanistic evidence that LLM emotion representations are structurally embedded along valence-arousal axes. SVAF forms Layer 4 of the Mesh Memory Protocol (MMP); the other half of the coupling engine is a per-agent Closed-form Continuous-time (CfC) neural network at Layer 6, whose learned per-neuron time constants (tau) create the temporal dynamics from which collective intelligence emerges: fast neurons synchronise affect across agents in seconds, while slow neurons preserve domain expertise indefinitely. SVAF determines what enters each agent's cognitive state; CfC determines how that state evolves. Trained on 237K samples from 273 narrative scenarios, SVAF achieves 78.7% three-class accuracy. We verify the complete mesh cognition loop -- from per-field evaluation through remix, CfC state evolution, tau-modulated peer blending, and autonomous action -- in a live deployment with 7 nodes across macOS, iOS, and web.

Figures & tables

Level	Question answered	Operates on	Mechanism
Peer-level	Is this agent cognitively close?	CfC hidden states	Cosine drift
Content-level (SVAF)	Is this signal relevant to me?	CMB fields	Per-field drift + fusion gate

Table 1

Field $f$	Axis	What it captures	Cross-Domain Relevance
focus	Subject	What the text is centrally about	Medium
issue	Tension	Risks, gaps, assumptions, open questions	Medium
intent	Goal	Desired change or purpose	Low
motivation	Why	Reasons, drivers, incentives	Medium
commitment	Promise	Who will do what, by when	Medium
perspective	Vantage	Whose viewpoint, situational context	Low

Table 1: The CAT7 field schema. Each CMB contains exactly these 7 fields.

State	Temp.	Trigger	Weight
observed	hot	Agent calls remember()	1.0 $\times$
remixed	warm	Peer remixes this CMB	1.5 $\times$
validated	warm	Human acts on this CMB	2.0 $\times$
canonical	cold	Validated + remixed by 2+ agents	3.0 $\times$
archived	whisper	No remix for 30 days	0.5 $\times$

Table 2: CMB lifecycle states. Anchor weight determines influence on future SVAF evaluations. Lifecycle progresses upward under activity (observed

\to

canonical); decays toward archived without it.

Agent	focus	issue	intent	motiv.	commit.	persp.	mood
Coding	2.0	1.5	1.5	1.0	1.2	1.0	0.8
Music	1.0	0.8	0.8	0.8	0.8	1.2	2.0
Fitness	1.5	1.5	1.0	1.5	1.0	1.0	2.0
Knowledge	2.0	1.5	1.5	1.0	0.5	1.5	0.3
Legal	2.0	2.0	1.5	1.0	2.0	1.5	0.5
Health	1.5	2.0	1.0	1.5	1.0	1.5	2.0

Table 3: Per-agent field weights (

\alpha_{f}

) across domains. Bold values indicate primary fields.

Agent	$\tau_{i}$	Rationale
Music agent	1,800s (30 min)	Current mood drives playlist
Coding assistant	7,200s (2 hr)	Current session context
Fitness agent	10,800s (3 hr)	Sedentary detection needs hours of context
Knowledge agent	86,400s (24 hr)	Daily digest cycle
Messaging agent	3,600s (1 hr)	Recent conversation context

Table 4: Per-agent freshness windows.

Component	Parameters	Purpose
FieldEncoder	7 heads $\times$ ( $128\to 64$ ) + backbone ( $384\to 256\to 128$ )	Learned per-field embeddings
CrossFieldAttention	4-head MHA + FFN	Field interactions within CMB
AnchorAttention	4-head MHA	Attend over local memory anchors per field
FusionGate	MLP ( $898\to 256\to 128\to 7$ )	Per-field gate values
FusionTransform	7 $\times$ MLP ( $128\to 64\to 64$ )	Non-linear per-field fusion
DriftPredictor	MLP per field + aggregate + decision	Drift scores + accept/reject

Table 5: SVAF model architecture.

Agent	Platform	Domain	Primary Fields	$\tau_{\text{freshness}}$
Coding assistant	macOS (Apple M4)	Coding assistance	focus, intent, issue	2 hr
Music agent	iOS (iPhone 14 Pro Max)	Music curation	mood, perspective, focus	30 min
Fitness agent	iOS (iPhone 14 Pro Max)	Fitness coaching	mood, issue, motivation	3 hr
Messaging agent	macOS (Apple M4)	Messaging	intent, mood, perspective	1 hr
Knowledge agent	macOS (Apple M4)	RSS aggregation	focus, issue, intent	24 hr

Table 6: Agent types used for training data generation.

Method	Aligned	Guarded	Rejected	Overall
Scalar (cosine, $\theta=0.5$ )	41.2%	79.3%	52.1%	66.8%
Scalar + temporal decay	43.5%	80.1%	55.8%	68.4%
Heuristic per-field ( $\alpha_{f}$ )	48.9%	83.4%	63.2%	73.1%
SVAF (neural)	57.8%	87.7%	70.5%	78.7%

Table 7: Classification accuracy by method and class.

Field	Mean $g_{f}$	Ratio to lowest	Cross-Domain Relevance
mood	0.497	8.9 $\times$	Fast — affect crosses all domains
focus	0.295	5.3 $\times$	Medium
issue	0.239	4.3 $\times$	Medium
commitment	0.121	2.2 $\times$	Medium
motivation	0.113	2.0 $\times$	Medium
intent	0.066	1.2 $\times$	Low

Table 8: Learned fusion gate values (mean over validation set).

Epoch	mood	focus	issue	commit.	motiv.	intent	persp.	3-class acc
1	0.331	0.006	0.005	0.009	0.008	0.011	0.005	72.8%
5	0.392	0.144	0.144	0.090	0.034	0.033	0.024	74.5%
10	0.369	0.129	0.111	0.108	0.020	0.017	0.034	77.2%
20	0.426	0.192	0.148	0.139	0.035	0.022	0.046	77.0%
30	0.490	0.248	0.206	0.135	0.099	0.053	0.047	78.5%
50	0.497	0.295	0.239	0.121	0.113	0.066	0.056	78.7%

Table 9: Gate evolution across training epochs.

Field	Extracted Text	Gate $g_{f}$	Effect
mood	“frustrated, low energy (v: -0.6, a: -0.4)”	0.50	Strongly accepted
focus	“coding with AI assistant”	0.30	Partially replaced by local context
issue	“exhausted, losing focus”	0.24	Accepted — fatigue is fitness-relevant
commitment	“coding session ongoing”	0.12	Mostly suppressed
motivation	“prevent burnout”	0.11	Partially accepted
intent	“needs break”	0.07	Mostly suppressed

Table 10: Per-field SVAF evaluation: coding

\to

fitness.

Parameter	Value
Model parameters	604,428
Training samples	188,480 (from 219 narratives, 20 agent types)
Validation samples	48,640 (from 54 held-out narratives)
Train/val split	By narrative (no narrative leakage)
GPU	NVIDIA A100-SXM4-80GB
Training time	2,454 seconds (40.9 min)

Table 11: Training hyperparameters.

Component	Latency	Notes
Heuristic SVAF (Node.js)	0.07 ms	Per-field drift, 1000 iter. averaged
Neural SVAF (Python, cold)	$\sim$ 5,800 ms	Includes PyTorch + model loading
Neural SVAF (persistent, warm)	$\sim$ 50–100 ms	Model loaded, inference only

Table 12: Inference latency comparison.

Agent	Platform	Role
COO	macOS (standalone node)	Cross-domain synthesis, pipeline, GitHub health
Research	macOS (standalone node)	arXiv papers, RSS feeds, paper monitoring
Marketing	macOS (standalone node)	Engagement tracking, content strategy
Product	macOS (standalone node)	PRD management, cross-product priorities
MeloTune	iPhone (iOS)	Music curation from mesh mood signals
MeloMove	iPhone (iOS)	Activity generation from mesh mood + local context

Table 14

Approach	What happens	Result
Message bus	A sends, B receives	B has A’s data. No new knowledge.
Pub/sub	A publishes, B subscribes	B has A’s data if on the right topic.
RAG	Agent retrieves similar docs	Agent has retrieved data. Single-agent.
MMP Remix	B processes A’s CMB via SVAF	New CMB that neither A nor B could produce alone. Graph grows.

Table 13: Remix vs. traditional data sharing.

Neuron type	$\tau$	Coupling behaviour	Role
Fast	$<5$ s	Couples readily	Mood, reactive signals
Medium	5–30s	Moderate	Context, activity patterns
Slow	$>30$ s	Resists coupling	Domain expertise — stays sovereign

Table 14: Per-neuron

\tau

-modulated blending in CfC.

Multi-Agent Collaboration

Multi-agent collaboration studies how autonomous agents coordinate, share information, and combine capabilities to solve tasks together.

See all 311 papers

Mesh Memory Protocol: Semantic Infrastructure for Multi-Agent LLM Systems
Apr 21, 2026Hongwei XuLLM Agent Memory Multi-Agent Communication
Superminds Test: Actively Evaluating Collective Intelligence of Agent Society via Probing Agents
Apr 24, 2026Xirui Li, Ming Li, Yunze Xiao +4Multi-Agent Coordination AI Agent Evaluation
University of Maryland · 2Mohamed bin Zayed University of Artificial Intelligence · 3Carnegie Mellon University
CoMem: Collective-Individual Memory Synergy for Evolutionary Multi-Agent Systems
Sep 14, 2026Chengxin Yu, Zhaoxin Fan, Faguo Wu +3Multi-Agent LLM Systems Multi-Agent Collaboration
Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing, School of Artificial Intelligence, Beihang University · Beijing academy of blockchain and edge computing · National University of Defense Technology +1

Apr 21, 2026J/K move · Enter open · S save

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

Hongwei Xu

LLM Agent Memory Multi-Agent Communication Agent Memory Management

Open paper AskPDF arXiv

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.

Abstract

Multi-Agent Collaboration

Explore similar work

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

Superminds Test: Actively Evaluating Collective Intelligence of Agent Society via Probing Agents

CoMem: Collective-Individual Memory Synergy for Evolutionary Multi-Agent Systems