Grimoire Cortex: Edge AI Orchestration

Status: Active Development (Internal Alpha)
Primary Discipline: Agentic Systems / Systems Engineering
Key Constraints: 8GB RAM Budget, Passive Cooling, Offline-First Execution

1. The Constraints (Physical Boundaries)

Grimoire Cortex is architected to thrive within the strict physical and thermal limits of edge hardware, specifically the Nvidia Jetson Orin Nano.

  • Memory: Hard cap of 8GB (unified memory). The system is tuned to keep the orchestration overhead under 1GB, maximizing the KV cache and parameter space for 4-bit and 8-bit GGUF models.
  • Thermal: Optimized for passive cooling. High-intensity loops are mitigated through asynchronous ritual dispatch to prevent thermal throttling in compact cyberdeck enclosures.
  • Connectivity: Zero-trust cloud reliance. All inference and tool execution occurs locally to ensure operational autonomy in field environments.

2. Architecture Decision Records (ADR)

ADR-001: Local GGUF over Cloud APIs

  • Context: Most modern agentic frameworks assume infinite cloud compute.
  • Decision: Standardize on native llama.cpp integration for C++ model execution.
  • Rationale: To achieve true agentic autonomy, the “Weight of the Node” must be physical. Local execution eliminates API latency, cost volatility, and data leakage.
  • Trade-off: Limits model selection to the 3B-8B parameter range to maintain interactive token velocity.

ADR-002: Asynchronous Ritual Execution

  • Context: Synchronous tool-calling blocks the inference engine, causing UI stutter and engine stalls.
  • Decision: Implement a decoupled “Ritual” pool for all I/O and hardware interaction.
  • Rationale: By treating tool execution as an asynchronous side effect (a “Ritual”), the engine can continue managing state and context while the hardware layer performs physical operations.

3. Hardware Bill of Materials (BOM)

Component Specification Purpose
Compute Nvidia Jetson Orin Nano (8GB) Primary AI Orchestration & Inference
Storage 1TB NVMe Gen4 Vector Store, Model Weights, & Event Logs
Power Omnicharge 20+ Portable DC Power & UPS
I/O 7” Waveshare DSI Display Real-time System Telemetry

4. Current Progress

The system currently bridges high-level intent parsing with low-level hardware triggers. Current work is focused on optimizing the context window preservation during multi-turn agentic loops on the Jetson’s unified memory architecture.

Development Logs

(Technical deep-dives and update logs will populate here as development progresses.)