The Modular, Self-Hosted Agentic Operating System

SanctumOS Architecture

SanctumOS is a self-hosted cognitive operating system for agentic AI. Its design is explicitly neuro-inspired, borrowing metaphors from human cognition while applying them to modular, auditable AI infrastructure.

Core Design Principles

  • Self-Hosting: All components run on personal hardware or small servers
  • Transparency: Every subsystem logs provenance for accountability
  • Extensibility: Plugin and modular design for easy extension
  • Neuro-Inspired: Subsystems mapped to cognitive functions
  • MCP-Native: The Model Context Protocol governs permissions and orchestration
  • Privacy by Default: No third-party centralized infrastructure

Cognitive Architecture

Thalamus

Function: Preprocessing hub for raw inputs

  • Phase 0: Immediate pass-through to Cerebellum
  • Phase 1: Cleanup, autocorrect, re-segmentation
  • Phase 2: Semantic annotation, topic tagging

Cerebellum

Role: Real-time filtering and prioritization

  • Low-latency event-driven filtering
  • Gut-check reflex arcs

Basal Ganglia

Role: Task selection and control loop management

  • Arbitration among competing sub-agent goals
  • Enforces success/fail loop structures

Conscious Mind

Core Agent: Central conversational and reasoning engine

  • Maintains core memory and recall memory
  • Coordinates sub-agents through MCP

Cochlea

Role: Audio input pipeline (research release)

  • Converts audio signals into structured input for the cognitive stack
  • Feeds into Thalamus, Broca, or Letta directly
  • Named after the inner ear structure that converts sound into neural signals

Dream Agent

Role: Nightly summarization and long-term integration

  • Hippocampus: Summarizes recent activity into knowledge candidates
  • Neocortex: Consolidates candidates into long-term structures
  • Glymphatic System: Compacts indexes and prunes superseded data

Infrastructure Layer

Sanctum Router ("Entorhinal Cortex")

Role: Self-hosted inference proxy

  • Routes agent LLM requests across multiple providers (Venice AI, Featherless, local servers)
  • Policy-enforced gating between agent identity and model inference
  • Credit-aware failover and deterministic routing

Sanctum Letta MCP (SMCP)

Role: Tool layer / Model Context Protocol server

  • Plugin-based architecture — each plugin exposes MCP tools
  • Auto-discovery of plugins at startup
  • Growing ecosystem of integrations (GitHub, Gmail, image analysis, cloud provisioning, etc.)

Broca-2

Role: Communication layer

  • CLI-first, plugin-based message processor
  • Translates agent decisions into platform-specific messages
  • Plugin system for endpoints (Telegram, CLI, web, APIs)
  • Message queue with echo, listen, live modes

Sanctum Web Chat

Role: Web interface bridge

  • Browser-based communication with agents
  • API-first design with admin monitoring

Memory Management

  • Core/Recall Memory: Active conversational state (Letta-managed)
  • Archival Memory: SQLite-backed persistent store
  • Summarization Cycles: Regular pruning and integration via Dream Agent
  • Infinite Context DB: Token usage tracking and context windows

Multi-Agent Ecosystem

  • Deep Research Agent: Structured analysis and synthesis
  • Internal Monologue Agent: Reflective insights generation
  • Dream Agent: Summarization and archival maintenance
  • Cerebellum: Input prioritization and filtering
  • Proactive Outreach Agents: Condition/time-based triggers

Development Philosophy

SanctumOS follows a neuro-inspired design that maps software components to cognitive functions, ensuring intuitive and predictable behavior that mirrors human cognitive processes.

Naming Convention

The Naming Rubric ensures consistent terminology across the system:

  • Neuroanatomical names for cognitive components (Thalamus, Basal Ganglia, Cochlea, Dream Agent)
  • Technical names for infrastructure (MCP, Router, Kernel)
  • Industry terms for AI/ML extensions (agents, tools, memory blocks)
  • Mythic names for agent personalities (Athena, Monday, Sentinel)

Next Steps

  1. Install: Use the Sanctum Installer (early alpha) or follow individual repo installation instructions
  2. Configure: Set up your first agent
  3. Explore Components: Start with the Components Overview or dive into Thalamus, Basal Ganglia, and Dream Agent
  4. Explore Modules: Learn about SMCP, Broca-2, and the SMCP Plugins
  5. Explore the Ecosystem: See reference applications built on the stack
  6. Extend: Build custom plugins

SanctumOS Architecture — a neuro-inspired approach to self-hosted AI agent systems.