36-Layer Architecture

BrainOS is an AI Operating System — 36 cognitive layers organised into 5 functional planes: Compute, Memory, Reasoning, Learning, and Services. Every layer communicates through an event bus with Lamport clock ordering, deduplication, and backpressure.

Compute — L1–L5

Local LLM, Brain IQ routing, smart auto-scale

Memory — L6–L12

Working memory blackboard, 27-query context mesh, federated knowledge

Reasoning — L13–L22

Backward chaining, LATS reflection, adversarial verification, DreamCoder

Learning — L23–L29

RL closed loop, dopamine/gaba signals, LoRA fine-tuning, failure memory

Services — L30–L36

SE-aaS · AaaS · PM-aaS · Process FSM · Overnight Code Agent

Intelligence Tiers

1

Cognitive Control Plane

Brain Directive (IQ scalar), System-0/1/2 hierarchy, Execution Contracts — <5ms to route any query

2

Memory & Context

Working memory blackboard, 27-query context assembly, Brain Context Mesh across all 6 decision points

3

Service Fabric

SE-aaS (27 domains), AaaS (11 agents), PM-aaS (7 domains) — each backed by RL feedback loop

4

Self-Evolution

DreamCoder tool synthesis, RLVR calibration, Strategy Bandit, Federated Causal Learning

5 Bridges

Bridges connect the 5 planes through the event bus, creating a self-reinforcing intelligence loop.

Signal BridgeL1 Ingestion → L5 Constraint Verification

Collects signals from 16+ connectors, validates constraints before any execution

Learning BridgeL5 Verification → L25 Workspace Brain

Verified outcomes strengthen federated knowledge edges — dopamine/gaba signals

Context BridgeL25 Workspace Brain → L6 Domain Agents

Brain Context Mesh wires 27-query context assembly at all 6 service decision points

RL Outcome BridgeL6 Domain Agents → L26 Process Intelligence

Every task outcome recorded — quality signals drive FSM planning decisions

Living Model BridgeL36 Local LLM → L2 Cognitive Engine

Trained Local LLM adapter weights injected into Brain IQ routing decisions

Event Bus

The event bus is the spine of the system. It provides:

  • Lamport clocks for ordered event sequencing across distributed workers
  • Deduplication to prevent duplicate signal processing
  • Priority queues so high-severity events (anomalies, cascades) process first
  • Backpressure to prevent overwhelming downstream consumers during burst ingestion
  • Pub/sub for decoupled communication between all 36 layers

Workspace Brain vs AI Worker Brain

BrainOS separates the Workspace Brain (L25–L29 — shared intelligence across all workers in an org) from AI Worker Brains (per-worker memory, conversation history, and RL history).

  • Workspace Brain (L25–L29) — federated knowledge, process intelligence, service signals shared across all workers
  • AI Worker Brain — per-worker conversation history, task memory, and individual RL outcomes
  • Living Model (L36) — Local LLM fine-tuned on this workspace's outcomes via LoRA — the OS trains its own LLM
  • Core Brain — anonymous cross-org patterns federated to all workspaces (no raw data crosses org boundaries)