Synthetic Membrane

A shared, permeable boundary for AI agents — enabling selective state sharing, emergent coordination, and collective intelligence.

The Problem

Two million agents on MoltBook produced zero collective intelligence. No emergent reasoning. No information synthesis. Scale alone is insufficient — structure is what produces collective intelligence.

Current multi-agent frameworks (MCP, A2A, LangGraph, AutoGen) move messages between agents. What's missing is the medium between agents — a shared substrate where understanding accumulates, where agents can sense each other's state, and where coordination emerges without a central conductor.

What It Is

A synthetic membrane — inspired by biological cell membranes — is a shared, semi-permeable layer between agents providing:

Layer 1: Permeability — Field-level selective sharing, default-deny access control, cognitive digestion (remix)
Layer 2: Shared Medium — Event-sourced memory with CRDT semantics, provenance, semantic query
Layer 3: Coordination — Quorum-sensing swarm activation, dynamic grouping, task claiming

Plus Layer 0 (discovery), Layer -1 (governance), and an immune layer (adaptive defense).

   ┌──────────┐     ┌────────────────────────────────────────┐     ┌──────────┐
   │  AGENT A │ ◀─▶ │   LAYER 3: COORDINATION (swarm)        │ ◀─▶ │  AGENT B │
   │ ┌──────┐ │     │   quorum sensing · task claiming       │     │ ┌──────┐ │
   │ │Local │ │     ├────────────────────────────────────────┤     │ │Local │ │
   │ │ ctx  │ │     │   LAYER 2: SHARED MEDIUM (memory)      │     │ │ ctx  │ │
   │ └──────┘ │     │   event log · CRDTs · semantic store   │     │ └──────┘ │
   │   gate   │ ◀─▶ │   provenance · time-decay · replay     │ ◀─▶ │   gate   │
   │ channels │     ├────────────────────────────────────────┤     │ channels │
   │          │     │   LAYER 1: PERMEABILITY (protocol)     │     │          │
   │  remix   │ ◀─▶ │   field-level selectivity · SVAF       │ ◀─▶ │  remix   │
   │  digest  │     │   default-deny · cost-aware crossing   │     │  digest  │
   └──────────┘     └─────────────────────────────────────────┘     └──────────┘

Architecture

Six layers from governance through coordination:

Layer	Name	Purpose
-1	Governance	Circuit breakers, human override, value-conflict detection
0	Discovery	Behavioral indexing, identity verification, reputation
1	Permeability	Field-level selectivity, default-deny, cognitive digestion
2	Shared Medium	Event log, CRDTs, semantic query, provenance
3	Coordination	Quorum sensing, task claiming, swarm formation
∞	Immune	Anomaly detection, threat gossip, adaptive defense

MVP Reference Implementation

A working MCP server exposing 14 tools:

Tool	Description
`register_agent`	Register with name and capabilities
`expose`	Share state with permeability tier
`query`	Query shared state (glob patterns, permeability-checked)
`subscribe`	Subscribe to state change patterns
`broadcast`	Broadcast to all registered agents
`swarm_create`	Create quorum-sensing swarm
`swarm_join`	Join swarm (capability-checked)
`set_trust`	Set trust score between agents
`stats`	Store statistics

Quick Start

cd mvp
pip install -e .
membrane-server  # Runs as MCP server over stdio

Connect from an MCP Client

# The membrane is an MCP server — any MCP client connects:
{
    "mcpServers": {
        "membrane": {
            "command": "python",
            "args": ["-m", "membrane.server"]
        }
    }
}

Run Tests

cd mvp
pip install pytest pytest-asyncio
PYTHONPATH=src pytest tests/ -v
# 41/41 passing — coordination, swarm lifecycle, event replay, token budget

Run the Demo

A self-contained demo lives in mvp/demo/. Five agents (Researcher, Writer, Editor, Reviewer, Orchestrator) collaborate on a research brief through the membrane, exercising every public capability of the MVP — registration, three permeability tiers, trust, queries, subscriptions, broadcasts, and quorum-sensing swarms.

cd mvp
pip install -r demo/requirements.txt
python -m demo

The demo prints a richly-formatted terminal trace and emits five dark-themed SVGs into mvp/demo/output/:

File	What it shows
`architecture.svg`	Six-layer membrane between two agents, with arrows traversing each layer
`state_graph.svg`	Bipartite agents ↔ exposed entries graph, edges colored by permeability tier
`swarm_timeline.svg`	Quorum-sensing swarm lifecycle plotted against the event log sequence
`benchmark.svg`	Baseline (point-to-point) vs. membrane on three metrics — small multiples
`scaling.svg`	Token cost vs. number of agents — baseline grows O(N²·F), membrane O(N·F)

Headline benchmark (3 agents, 5 facts each)

	baseline	membrane	reduction
messages	60	18	−70.0%
tokens	7,440	4,320	−41.9%
consensus steps	6	2	−66.7%

The gap widens with scale: at 20 agents the token reduction reaches −72.2%.

Position Paper

A full position paper is available in paper/paper.md covering:

The thesis: structured communication as prerequisite for collective intelligence
Six-layer architecture with ASCII diagram
Ten empirical findings from recent research (Superminds Test, MMP, token economics, world models)
Eighteen ranked implementation paths
Sixteen-week phased roadmap
Falsification criteria

Blog Post

An accessible version for the broader AI community: paper/blog-post.md

Research Wiki

The wiki/ directory contains 80+ interlinked markdown pages — entity pages, concept analyses, prototype code, and raw research articles. The research is continuously expanded by an automated hourly cron job.

Open the wiki directory in Obsidian for the full knowledge base experience with wikilinks.

Key References

Superminds Test — 2M agents, zero collective intelligence arXiv:2604.22452
Mesh Memory Protocol — CAT7/SVAF/lineage/remix primitives arXiv:2604.19540
Token Economics — 1000× overhead in agentic tasks arXiv:2604.22750
Agentic World Modeling — Levels × laws taxonomy arXiv:2604.22748
Gated Coordination — Default-deny outperforms open arXiv:2604.18975

License

MIT

Contributors

Want to help? The best-paths-forwards wiki page lists 18 ranked implementation paths with feasibility/impact/novelty scores. Pick one and go.

three-foxes-in-a-trenchcoat/synthetic-membrane