Personal AI Ecosystem

Geoffrey (Personal AI Assistant) & Open Brain

A personal AI ecosystem where every interaction feeds a shared knowledge base, creating a self-improving context loop. All code generated using Claude Code (Opus).

Multi-channel access Telegram, SMS, phone, web, Siri, Slack, Apple Watch

Persistent memory pgvector semantic search + auto-tagging

Self-improving loop Every interaction adds context

~$5/month infrastructure Free tiers + Hetzner VPS

System Components

Three AIs, One Brain

Geoffrey handles daily assistance, Matthew handles reflection, and Open Brain is the persistent memory layer they share.

🤵

Geoffrey

Personal AI Assistant — Modelled after Alfred

A personal AI assistant available across every major channel. Handles web search, calendar, email drafts, and routes calls to other AI assistants.

Real-time web search (Brave Search API)
Persistent memory via Open Brain
Proactive notifications & check-ins
Calendar read + email drafts (Telegram-gated)
Call transfers to other assistants
Two system prompts: voice (Vapi) and text (n8n)

🧠

Matthew

Reflection Coach AI

Triggers end-of-day reflection calls via geofencing. Nordic-style directness with a structured two-step response pattern.

End-of-day calls triggered by geofencing
Nordic-style directness
Two-step response pattern
Anti-parroting rules
Captures reflections at end of call
Tracks date, mood, energy, events, successes, frustrations, relationships, reflections, intentions

💾

Open Brain

Personal Knowledge System

Full CRUD personal knowledge store with hybrid retrieval. Extended from Nate B. Jones' open-source concept.

Full CRUD (edit and delete, not just capture)
Hybrid retrieval (semantic + chronological + tag)
Automatic topic tagging + people extraction
Auto-splitting compound thoughts
5 thought types: observation, task, idea, reference, person_note
Thought statistics + web search augmentation

Architecture

MCP Servers & Data Layer

Contact channels flow through n8n into MCP servers, which interface with Supabase Postgres + pgvector.

Contact Channels

✈️ Telegram

💬 SMS

📞 Phone / Vapi

🌐 Web / Next.js

🍎 Siri / iOS Shortcuts

⌚ Apple Watch

MCP Servers

open-brain-mcp
Tool	Purpose
`capture_thought`	Save a new thought
`search_thoughts`	Semantic + tag search
`list_thoughts`	Chronological listing
`edit_thought`	Update existing thought
`delete_thought`	Remove a thought
`thought_stats`	Usage statistics

web-search-mcp
Tool	Purpose
`realtime_search`	Brave Answers API with security validation

ingest-thought
Endpoint	Purpose
`POST /ingest`	Slack webhook for capturing thoughts

Built with @modelcontextprotocol/sdk, Hono, Zod, streamable HTTP transport.
Authenticated via MCP_ACCESS_KEY.

Database

Supabase Postgres + pgvector

thoughts table

iduuid

contenttext

embeddingvector(1536)

metadatajsonb

created_attimestamptz

RPC Function

match_thoughtsrpc

LLM Usage — via OpenRouter

Task	Model
Embeddings	openai/text-embedding-3-small
Metadata extraction	openai/gpt-4o-mini
Thought splitting	openai/gpt-4o-mini
Tag extraction for search	openai/gpt-4o-mini

The Flywheel

Self-Improving Context Loop

Each AI interaction feeds the knowledge base. The more context Open Brain accumulates, the smarter every AI response becomes.

Matthew calls Tim

end-of-day reflection

→

Extracts reflections

mood, energy, events

Geoffrey searches knowledge

context-aware responses

←

Embeds, tags & indexes

pgvector + metadata

Tim captures thoughts

Siri, Telegram, Slack

Saves to Open Brain

capture_thought MCP

Every interaction adds context. Every AI tool gets smarter.

Infrastructure

Tech Stack

Designed for near-zero cost. Free tiers cover everything except a single VPS.

Technology	Role
Supabase	Postgres + pgvector + Edge Functions hosting MCP servers
n8n	Central workflow automation — channel routing, API orchestration
Twilio	SMS routing + phone call routing to Vapi.ai
Vapi.ai	Realtime voice — STT/TTS for phone and web calls
OpenRouter	LLM routing — abstraction layer for model selection
Telegram Bot API	Primary chat interface
Brave Search API	Real-time web search via custom MCP server
Claude Code (Opus)	All development — no manual code written
Vercel	Next.js web interface (free tier)
Hetzner VPS	Docker hosting for self-hosted n8n (~$5/month)
Cloudflare	DNS + DDoS protection
GitHub	Source code repository

~$5/month total infrastructure cost

Free tiers cover Supabase, Vercel, Cloudflare, and GitHub. The only paid component is a Hetzner VPS for Docker-hosted n8n.

Domain Structure

Subdomain	Target	Cloudflare
n8n.digitomnia.com	Hetzner VPS	Proxied (DDoS protection)
mcp.digitomnia.com	Supabase Edge Functions	DNS only
dash.digitomnia.com	Vercel	DNS only

Capture Interfaces

Every Context, Every Surface

Open Brain accepts thoughts from any channel — hands-free, chat, voice, or developer tooling.

🗣️

Siri

iOS Shortcuts integration for hands-free capture

💬

Via Geoffrey — the primary chat interface

📞

Voice Calls

Via Matthew's daily reflection calls

🌐

Web Interface

Next.js dashboard with voice + chat

📌

Slack

Via ingest-thought webhook

🤖

MCP Clients

Any MCP-compatible AI (Claude, ChatGPT, etc.)