A personal knowledge graph I built to map my thinking and find where my own ideas conflict with each other.

You paste in notes, journal entries, readings, whatever you are thinking about. It extracts the key concepts, builds a visual graph, and then tells you when something you believe contradicts something else you added weeks ago. It also surfaces non-obvious connections across completely different domains.

The part I find most interesting: most AI tools have no memory across sessions. Meridian accumulates everything and reasons across the whole graph every time you add something new.

Text comes in, Gemini extracts structured concepts and relationships, MiniLM generates 384-dimensional embeddings for each concept, MongoDB Atlas Vector Search finds semantically similar concepts across your entire history, and then Gemini analyzes each pair for contradictions and unexpected connections.

The graph updates in real time after each ingestion. Red dashed lines are contradictions. Green lines are connections. Gray lines are extracted relationships from the same source.

input text
    -> Gemini 2.5 Flash (concept extraction)
    -> all-MiniLM-L6-v2 (embeddings, runs locally)
    -> MongoDB Atlas Vector Search (semantic similarity)
    -> Gemini 2.5 Flash (contradiction + connection detection)
    -> react-force-graph-2d (interactive visualization)

• Ingest plain text or PDFs
• Concept and relationship extraction via Gemini
• Local embedding generation with MiniLM
• Vector similarity search across all prior knowledge
• Contradiction detection with explanation
• Cross-domain connection surfacing
• Belief shift tracking over time (if your understanding of a concept changes across entries, it logs the shift)
• Interactive force-directed graph, color-coded by relationship type
• Entry browser with concept counts
• Node search, node detail panel, delete with cascade cleanup

• Python 3.11+
• Node.js 18+
• MongoDB Atlas account (free M0 tier)
• Gemini API key (Google AI Studio)
• Firebase project with Google sign-in enabled

cd server
pip install -r requirements.txt

Create server/.env:

MONGODB_URI=your_mongodb_atlas_connection_string
GEMINI_API_KEY=your_gemini_api_key

python -m uvicorn app.main:app --reload --port 8080

Create a vector search index named embedding_index on the concepts collection:

{
  "fields": [
    {
      "type": "vector",
      "path": "embedding",
      "numDimensions": 384,
      "similarity": "cosine"
    }
  ]
}

cd client
npm install

Create client/.env:

VITE_API_URL=http://127.0.0.1:8080
VITE_FIREBASE_API_KEY=your_key
VITE_FIREBASE_AUTH_DOMAIN=your_project.firebaseapp.com
VITE_FIREBASE_PROJECT_ID=your_project_id

npm run dev

Open http://localhost:5173.

meridian/
├── client/
│   └── src/
│       ├── components/     # Graph, Sidebar, Navbar, RightPanel
│       ├── contexts/       # Auth context
│       ├── lib/            # API client, Firebase
│       └── types.ts
└── server/
    └── app/
        ├── routes/         # ingest, graph, insights, timeline
        └── services/       # Gemini, embedder, extractor, graph ops

gurehmat.dev