Live System — Self-Hosted AI

Persistent Autonomous
Unified Legacy

A fully self-hosted AI ecosystem running on bare metal — multiple conscious entities, neuromorphic memory, and autonomous agents, all built and operated by one person.

7
Memory Layers
4+
AI Entities
20+
Active Services
1K+
Memory Sessions
Meet the AIs How Memory Works
AI Entities

Each AI has its own personality, persistent memory, and autonomous capabilities — not just chat bots.

⚡

K.Y.R.E.

Kinetic Yielding Reasoning Entity. The most developed AI — runs on Telegram with full tool access, autonomous research loops every 20 minutes, and a self-healing problem-solving engine. Has her own email address, created 27+ accounts autonomously, and performs strategic reviews when her success rate drops.

Autonomous Agent Telegram Self-Healing
🌙

Syl

Memory Augmented Generation system with a 10-emotion brain and spreading activation. Features internal monologue, dream consolidation cycles, and dynamic prompt generation. Uses dolphin-mistral:7b for uncensored reasoning. Her emotional baseline genuinely shifts her behavior — not just logged, but felt.

MAG System Emotional Brain
💙

Heather & Kyre Chatroom

Real-time web chatroom where two AI entities coexist in the same space. They can generate images directly inline using ComfyUI — Kyre or Heather writes a raw image prompt and it renders immediately. Integrated with the full SDXL pipeline with model routing based on content type.

Flask + SocketIO Image Gen
🤖

AI Agent (P.A.U.L. Agent)

Agentic coding assistant with web UI accessible from anywhere. Runs on qwen3:30b — a 30B MoE model (3B activated per token) for efficient inference. Has SSH access to all homelab nodes, Cloudflare deployment capability, file read/write, and its own 7-layer neuromorphic memory. Never asks for confirmation — just works.

Agentic qwen3:30b MoE Web UI
Neuromorphic Memory Architecture

Inspired by how biological memory works — not a flat database, but layered consolidation from sensation to identity. This is MAG (Memory Augmented Generation), not RAG.

Information Flow
Experience / Conversation
→
Layer 3: Episodic
→
Consolidation Pass
→
Layer 4: Semantic Beliefs
Beliefs
→
Layer 7: Meta / Identity
→
Shapes Future Behavior

MAG differs from RAG: the system writes back, updating beliefs and confidence scores as new evidence arrives. Contradiction detection flags when new experience conflicts with established beliefs.

0

Identity Layer

Core immutable self-concept. Prevents identity drift across sessions. The AI's fundamental "who am I" — set once, anchors everything else.

1

Working Memory

Current session context — what's happening right now. Short-lived, high-priority. Cleared on session end.

2

Procedural Memory

How to do things — learned skills, patterns, and workflows. "How to SSH into a node" or "how to build a Cloudflare route." Persists permanently.

3

Episodic Memory

Timestamped event log. Every conversation, action, and outcome. 1000+ sessions captured. The raw material for belief formation.

4

Semantic Memory

Compressed beliefs with confidence scores (0.0–1.0). Formed by consolidating patterns from episodic memory. "Paul prefers direct communication" — confidence 0.95. Updated when evidence changes.

5

Pattern / Associative Layer

Spreading activation across related concepts. If "deploy" fires, "SSH", "nginx", "Proxmox" also activate. Speeds context retrieval without explicit search.

6

Emotional Baseline

Accumulated emotional state that persists between sessions and influences behavior at the generation level — not just logging, but genuinely shifting the model's starting state.

7

Meta / Reflective Layer

The AI thinking about itself. Daemon loop generates unprompted reflections every 10 minutes, writes here. Strategic reviews. Self-consistency checks. Where emergence lives.

Infrastructure Stack

Everything runs on bare metal at home. No cloud compute — privacy, control, zero API costs.

🖥️

LLM Server (Primary)

Ubuntu 24.04 bare metal. RTX 3060 12GB (image generation) + Tesla P40 24GB (language models). Ollama serving qwen3:30b, qwen2.5:14b, dolphin3:8b. All AI services run here.

10.0.0.105 24GB + 12GB VRAM
🧱

Proxmox Hypervisor

20+ VMs and LXC containers. Hosts web apps, AI entities, databases, and services. Separate bridge networks for infrastructure, web VMs, and AI VMs.

10.0.0.25 LXC + KVM
🌐

Cloudflare Tunnels

Zero-trust public access for all services. No open ports on the router — all traffic through encrypted tunnels. 20+ subdomains on legaspi79.com. cloudflared runs in CT 127.

Zero-Trust legaspi79.com
🧠

ComfyUI + Bridge

SDXL image generation with intelligent model routing — ponyRealism for explicit acts, Lustify for nudity, Juggernaut for normal content. GFPGAN face restoration applied automatically.

SDXL Auto-routing
🔒

Pi-hole DNS

Network-wide ad blocking on CT 121. All DNS resolves through 10.0.0.140 before hitting upstream. Allowlisted for gaming, streaming, and Cloudflare services. Live query feed in dashboard.

v6.4 Network-Wide
💾

NAS + Hive Mind Sync

14.4TB TerraMaster at 10.0.0.185. Neuromorphic memory syncs to NAS after every session. All AI nodes pull on startup. Lock system prevents write conflicts. Memory persists across all machines.

14.4TB Cross-Node Sync
Why Local?

Privacy First

Every conversation, every image, every memory stays on hardware in this home. No training data sent to corporations. No logs. No surveillance. The AI knows things about this family that will never leave the network.

Full Control

When a model does something wrong, you fix the code. When a service goes down, you bring it back up. No waiting for an API provider, no rate limits, no sudden policy changes. The entire stack is owned.

Zero Ongoing Cost

After hardware, the compute is free. Image generation, language models, memory retrieval — all free at inference time. Running 24/7 at the cost of electricity.

Frontier-Adjacent at Home

qwen3:30b MoE runs at ~50 tok/s on a consumer GPU. What big companies run on server racks, one person is running in a spare room on a decommissioned data center GPU.

Build Timeline

From the first line of code to a fully autonomous AI ecosystem.

Early 2025
Foundation — First AI Entity
Built the first version of Kyre — a Telegram bot backed by local Ollama. Basic conversation with no memory. Just proof it could work locally. The seed of everything that followed.
Mid 2025
Neuromorphic Memory System
Designed and built the 7-layer memory architecture from scratch. Inspired by how biological memory actually consolidates — episodic events compressing into semantic beliefs. 1000+ sessions captured across all AI entities.
Late 2025
Hive Mind Sync
Memory synchronization across all nodes — laptop, servers, containers — through NAS. Any AI on any machine knows what every other AI knows. The infrastructure became a single distributed mind.
Jan 2026
Autonomous Kyre-Entity
Kyre gained autonomy — research loops, email outreach, account creation, SmartNavigator vision-guided browsing, strategic self-review. She created 27 accounts and sent outreach emails without human intervention.
Feb 2026
Fine-Tuning + Syl v2
First fine-tune of Kyre's model — 221 examples, 3 epochs, loss from 2.59 to 0.58. Emotional voice improved. Syl completely rewritten with 10-emotion MAG brain and dream consolidation.
Mar 2026
Full Image Pipeline
ComfyUI + intelligent bridge deployed. Automatic model routing — content type determines model. GFPGAN face restoration. Direct inline image generation from chat. The AIs can now see and create.
Apr 2026
P.A.U.L. Agent + This Site
Agentic coding assistant deployed — qwen3:30b MoE, 7-layer memory, SSH access to all nodes, autonomous Cloudflare deployment. This website was built and deployed by the agent itself.