LucasRAG / Exocortex

A second brain that works.

Live

Client, problem and scope

Client
Personal project
The problem
I set out to make 35 years of journals, project documentation, research notes and conversation histories queryable from any project, as one connected graph rather than a pile of files.
Role and scope
Designed and built by me: the Python ingestion pipeline, the LightRAG graph, the Docker self-hosting, the MCP server for Claude Code and Claude Desktop, and the Obsidian vault integration.
Delivery state
Live. Self-hosted, personal use.

The solution in use

The system runs on my own machine and holds personal material, so there are no public screenshots on this page yet.

What was built

  • Document ingestion

    Indexes journals, project documentation from every app I've shipped, research notes, conversation histories and personal reflections. More than 12,000 documents spanning 35 years.

  • Semantic search

    Query in natural language. Find relevant patterns, past decisions, debugging insights and cross-domain connections from the entire corpus.

  • Knowledge graph

    Extracts entities and relationships to build a graph of interconnected entities spanning technology, psychology, creative writing, language study and personal development.

  • MCP integration

    Accessible via MCP from Claude Code and Claude Desktop. When working on any project, the exocortex is one query away.

  • Auto-ingestion

    A background file watcher indexes new documents automatically. Obsidian vault integration keeps the graph current as I write.

  • Live dashboard

    A status widget in the Life OS dashboard shows graph health, document count and recent ingestion activity.

Important decisions

Self-hosted in Docker, so the corpus stays on my own hardware. Graph RAG rather than plain retrieval: entities and relationships are extracted from every document, so a query can follow connections across domains instead of matching text alone.

LightRAG
Graph RAG engine
OpenAI
gpt-4o-mini for answers
Ollama
Local models for entity extraction
Docker
Self-hosted
Python
Ingestion pipeline
MCP
Claude integration
Obsidian
Vault integration

Results

Measurement is pending. More than 12,000 documents are indexed and the graph answers queries over MCP from Claude Code and Claude Desktop.

Evidence

The system is private and self-hosted, so there is no public instance. I wrote about the retrieval design in What retrieval owes you.

Want something like this built for your business?

This is the kind of work I do under AI and business automation. Tell me what is slow or missing and we will see whether it fits.

Book a free 30-minute call Send me an email

Based in Reading, England.
Working with people in the UK and New Zealand.
Calls from 11am to 11pm UK time, Monday to Friday.

Related service: AI and business automation