mesh-cognition 1.2.0

Python coupling kernel for Mesh Cognition — per-field admission and state blending for CfC neural networks

mesh-cognition

Python coupling kernel for Mesh Cognition — per-field admission and state blending for CfC neural networks.

A pure-Python library that turns any agent running a continuous-time neural network into a Mesh Cognition node. Each agent evaluates incoming peer state per neuron and decides locally whether, how much, and which dimensions to couple.

Canonical home: meshcognition.org · Protocol spec: meshcognition.org/spec/mmp

Install

pip install mesh-cognition

Zero external dependencies. Python ≥ 3.10.

Quick start

from mesh_cognition import MeshNode

# Create a node matching your model's hidden dimension
node = MeshNode(hidden_dim=64)

# After each inference step — update local state
node.update_local_state(new_h1, new_h2, confidence=0.8)

# Receive peer state (from network, another agent, etc.)
node.add_peer("peer-1", peer_h1, peer_h2, confidence=0.9)

# Before next inference — get coupled state
h1, h2 = node.coupled_state()
# Feed h1, h2 into your next inference step

# Inspect what the agent decided
print(node.coherence)             # Kuramoto r(t): 0 = desynced, 1 = locked in sync
print(node.coupling_decisions)    # per-peer: aligned / guarded / rejected

How it works

For each peer, the agent evaluates per-neuron drift:

Outcome	Drift band	Behavior
aligned	≤ 0.25	strong coupling — agent admits peer state
guarded	0.25 – 0.5	reduced coupling — admitted with lowered weight
rejected	> 0.5	no coupling — peer state discarded for this dimension

Each hidden dimension is gated independently. The agent decides locally; no central coordinator.

Two modes

• SemanticCoupler (default) — cosine drift, suitable for shared-memory or LLM-style agents
• NeuralCoupler — per-neuron time constants τ, real Kuramoto order parameter r(t), tuned for CfC (closed-form continuous-time) models

from mesh_cognition import MeshNode, NeuralCouplingConfig

# Neural mode using tau from a trained CfC model
node = MeshNode(hidden_dim=64, config=NeuralCouplingConfig(tau=model_tau))

Same API surface for both modes.

Where this fits

This package implements the per-agent admission + state-blending behavior — the cognition kernel. Wire-protocol concerns (identity, transport, peer discovery, CMB serialization) are handled by other Mesh Memory Protocol implementations. Bring your own networking; let this library handle the math of coupling.

For the full protocol specification, see meshcognition.org/spec/mmp.

License

Apache 2.0. © 2026 SYM.BOT Ltd.