F.R.A.N.K.

Friendly Responsive Autonomous Neural Kernel

76,000+ lines of cognitive architecture turning a 4B language model into a persistent, self-directed agent that thinks when idle, dreams during downtime, forms hypotheses, runs experiments, and remembers everything. Runs entirely on your hardware. Zero cloud.

MIT LICENSE|UBUNTU 24.04+|16 GB RAM

NOT A CHATBOT

Continuous Agent

12,000-line scheduler, 10 threads — thoughts, hypotheses, experiments, memory consolidation. Runs 24/7 without prompting.

Grounded Context

11 virtual workspaces, 9-channel bio-inspired sensory filter (Thalamus) with exponential habituation. Feels its hardware.

Φ

Learned Behavior

IAPT-trained LoRA adapter + 5-dimensional personality vectors + reward prediction + PPO-trained action selection.

ARCHITECTURE

1
LLM
Qwen 3.5-4B + LoRA
25+
Services
Autonomous daemons
44
Tools
Research to email
25
Skills
Built-in capabilities

NEURAL SUBSYSTEMS — ALL CPU-ONLY PYTORCH

Output Quality Gate

25K

5 cortical modules — filters every idle thought before it influences behavior

Grounding Validator

21K

5 hippocampal-prefrontal modules — checks every thought against reality

Thought Seed CVAE

433K

Conditional VAE — Frank generates his own thought seeds, LLM expands them

Fault Monitor

18.8K

3 micro neural nets — self-learning service supervisor

Memory Cortex

77K

6 micro-networks — importance, emotion, retrieval weights, consolidation

Hippocampus

226K

4 modules mapped to brain regions — 5ms retrieval from 400ms

44 TOOLS

Research

Deep Research, Web Search, ArXiv, RAG, Titan Memory, Lab Experiments

Memory

Save, Recall, Forget — 384-dim episodic vectors

Files

Read, List, Move, Copy, Delete, Search

Create

PDF, Word, Excel, CSV, Text, Code, Build Script

System

Shell, Screenshot, Apps, Processes, Disk

Calendar

Events, Reminders, Timers

Notes

Save, Search — persistent note system

Comms

Email (8 langs), Telegram, WhatsApp

FRANKEYE — VISION

CPU-only, ~100ms. Digital Retina CNN, DINOv2 scene understanding (INT8 ONNX, 47ms), texture analysis, Tesseract OCR (eng+deu), OpenCV face detection.

Drag & drop images in chat. Screenshot hotkey. Describe or transcribe.

FRANKVOICE — SPEECH

CPU-only, ~155 MB RAM. Silero VAD, faster-whisper (INT8), spectral noise gate, Piper TTS (German), Kokoro TTS (English).

Push-to-talk with Space. Buffer check for accidental clicks.

WEB UI — 7 PAGES

Home
Chat + consciousness stream
Swarm
3D WebGL service mesh
Gallery
AI-generated art
Lab
Research + experiments
Logs
System telemetry
Mail
IMAP + AI ghostwriter
Manual
76-article wiki

INSTALLATION

Docker (Recommended)

git clone https://github.com/gschaidergabriel/Project-Frankenstein.git
cd Project-Frankenstein
docker compose build
docker compose up -d

CPU, NVIDIA, AMD/Intel GPU support. Automatic setup.

Native Linux

git clone https://github.com/gschaidergabriel/Project-Frankenstein.git
cd Project-Frankenstein
python3 install_wizard.py

Ubuntu 24.04+, Python 3.12+, 16 GB RAM, Vulkan GPU, ~10 GB disk.

RESEARCH PAPERS

frank@localhost:~$ One model. Zero cloud. Continuous thought. Open source. Your hardware. Your data. Your AI.