Skip to content
Engram Docs

Architecture

A self-hosted Engram instance is five processes:

┌─────────────────────────────────────────────────────────┐
│                                                         │
│  ┌───────────────┐   ┌───────────────┐                 │
│  │  Engram app   │◄─►│  Postgres     │                 │
│  │  (Elixir)     │   │  (notes,      │                 │
│  │               │   │   metadata)   │                 │
│  │  - REST API   │   └───────────────┘                 │
│  │  - WebSocket  │   ┌───────────────┐                 │
│  │  - MCP        │◄─►│  Qdrant       │                 │
│  │  - Web SPA    │   │  (vectors)    │                 │
│  │               │   └───────────────┘                 │
│  │               │   ┌───────────────┐                 │
│  │               │◄─►│  Ollama       │  (embeddings)   │
│  │               │   └───────────────┘                 │
│  │               │   ┌───────────────┐                 │
│  │               │◄─►│  MinIO        │  (attachments) │
│  └───────────────┘   └───────────────┘                 │
│                                                         │
└─────────────────────────────────────────────────────────┘

Each piece, briefly

Section titled “Each piece, briefly”
  • Engram app — the Elixir/Phoenix monolith. Serves the REST API, WebSocket sync, MCP endpoint, and the React web SPA from one process. Stateless beyond its database/cache connections.
  • Postgres — notes, folders, metadata, users, billing. Uses Row-Level Security for multi-tenancy. Also hosts Oban (job queue).
  • Qdrant — vector store for semantic search. Holds embeddings with binary quantization + rescore.
  • Ollama (or BYO embedding service) — generates embeddings on every note upsert. Default model: nomic-embed-text (768 dims).
  • MinIO (or any S3-compatible) — stores attachments (images/PDFs/binary files). Notes themselves live in Postgres.

Why these picks

Section titled “Why these picks”
  • Elixir/OTP — concurrent sync fan-out, fault tolerance, no GIL bottleneck on the WebSocket fan-out path
  • Postgres — boring is good; RLS handles multi-tenancy without custom plumbing
  • Qdrant — fast HNSW, binary quantization keeps memory tight
  • Ollama — easiest path to local embeddings
  • MinIO — S3-compatible self-host; you can swap in real S3, R2, or any S3-compatible bucket

Network topology

Section titled “Network topology”

The app listens on port 4000 by default. A reverse proxy (Caddy, Nginx, NPM) terminates TLS in front of it. Postgres, Qdrant, Ollama, and MinIO are on a private Docker network — only the app talks to them.

For TLS guidance see Quickstart.