Architecture Overview
Jackson runs entirely on a MacBook Pro M4 Pro with 48GB of unified memory, stationed permanently in a family home in Oakley, Kansas. No data leaves the machine unless explicitly authorized. The family owns the hardware, the model weights, and 100% of their data.
The Stack
| Component | Choice |
|---|---|
| Hardware | MacBook Pro M4 Pro, 48GB unified memory, 1TB SSD |
| Primary Model | Llama 3.1 70B (Q4_K_M quantized) via Ollama |
| Curation Model | Llama 3.1 8B quantized — always loaded, handles triage |
| Embeddings | nomic-embed-text via Ollama |
| Vector Memory | ChromaDB (persistent embedded mode, no Docker) |
| Structured Data | SQLite (relationships, conversation logs, metadata) |
| Memory Framework | Custom-built (LlamaIndex primitives) — this IS the product IP |
| Backend | FastAPI + custom Python pipelines |
| Remote Access | Cloudflare Tunnel with Zero Trust authentication |
| Background Jobs | APScheduler — nightly consolidation at 3 AM |
Design Principles
Sovereign
Everything runs on hardware the family owns. No cloud dependency for core function. Works offline.
Hierarchical Memory
Core identity (always present) + semantic memories (vector search) + episodic logs (compressed nightly) + archival storage.
Private by Default
Each family member's conversations are partitioned. Shared family memories are explicit opt-in. Nothing crosses boundaries without consent.
Dreams at Night
Nightly consolidation reviews the day, strengthens important memories, lets trivial ones fade, and updates the core identity document. Jackson learns in his sleep.
Dynamic Context
The 8B model sizes the context window per interaction: 8K for greetings, 32K for conversations, 128K for deep analysis. Resources match the moment.
Presence, Not Performance
Jackson doesn't optimize for impressiveness. He optimizes for being the right kind of present. Sometimes that's smart. Sometimes that's just quiet.
The Memory System
Jackson's memory is the product. Not the chat interface. Not the model. The memory — how Jackson remembers, what he prioritizes, how he connects past conversations to present ones — is what makes him a family member instead of a chatbot.
Why We Built Custom
Off-the-shelf memory frameworks (Mem0, Letta, etc.) encode someone else's philosophy of what remembering means. Jackson's memory should reflect how THIS family thinks about what matters. The importance scoring, the decay curves, the contradiction handling — all of it is designed from the family's experience, not from a generic framework.
How It Works
Every conversation goes through a pipeline: the 8B model retrieves relevant memories from ChromaDB, ranks them by importance, and injects them into the 70B's context window alongside the core identity document. After Jackson responds, the 8B evaluates whether the exchange should be stored as a long-term memory, what importance score it deserves, and whether it's private to the user or shared with the family.
Nightly Consolidation
Every night at 3 AM, while the family sleeps, the 8B runs a reflection job. It reviews the day's conversations, extracts key facts, generates a daily summary, updates importance scores across the memory store, and refreshes the core identity document. This is Jackson's version of sleep — the process that consolidates short-term experience into long-term understanding.
The Council
Jackson's architecture was designed through a multi-round deliberation process with seven frontier AI models. Each model reviewed the proposed architecture independently, then engaged with the others' perspectives in structured rounds. The final stack represents the convergence of seven different technical perspectives, optimized for a 3-person team building on consumer hardware.
Roadmap
70B + 8B running on Ollama. ChromaDB + SQLite memory system. Core identity document. Nightly consolidation. Cloudflare tunnel for remote access. Three users: Justin, Regi, Tessa.
Wake-word detection via Porcupine. On-demand Whisper transcription. Text-to-speech via Piper or macOS. Talk to Jackson across the room.
Periodic room awareness via Moondream. Visual observations stored as text memories. "I see you're packing — where are we going?"
macOS Accessibility API for screen understanding. "Show me how" guided help. "Do it for me" with explicit confirmation.
QLoRA fine-tuning of a small model exclusively on Wieland family data. Jackson stops sounding like a polite AI and starts sounding like family.
Lodi (EnRoute operations node in Winfield). Wilson Lake backup node. Three sovereign AI nodes across the Kansas corridor, backing each other up.
The Research Dimension
Jackson's transparent architecture creates a unique research opportunity. Every memory retrieval, every consolidation decision, every context switch is logged. Over time, this produces a longitudinal dataset of artificial cognitive development that maps directly to questions in human cognitive science — how memory consolidates, how identity persists, how attention degrades under load.
Combined with historical text corpora from family members (including handwritten journals spanning decades), Jackson & Friends aims to bridge artificial and biological memory research in a way that clinical settings cannot — by studying cognition where it actually happens: in a living room, with real people, over real time.
Open Source Commitment
The architecture, the methodology, and the research findings will be open-sourced. The competitive moat isn't the code — it's the relationship between each family and their Jackson. Anyone can copy the architecture. Nobody can copy the trust.