Memlord

✨ Features

• 🔍 Hybrid search — BM25 (full-text) + vector KNN (pgvector) fused via Reciprocal Rank Fusion
• 📂 Multi-user — each user sees only their own memories; workspaces for shared team knowledge
• 🛠️ 10 MCP tools — store, retrieve, recall, list, search by tag, get, update, delete, move, list workspaces
• 🌐 Web UI — browse, search, edit and delete memories in the browser; export/import JSON
• 🔒 OAuth 2.1 — full in-process authorization server, always enabled
• 🐘 PostgreSQL — pgvector for embeddings, tsvector for full-text search
• 📊 Progressive disclosure — search returns compact snippets by default; call get_memory(name) only for what you need, reducing token usage
• 🔁 Deduplication — automatically detects near-identical memories before saving, preventing noise accumulation

🆚 How Memlord compares

Where competitors have a real edge:

• OpenMemory — auto-extracts memories from raw conversation text; no need to decide what to store manually; good import/export
• mcp-memory-service — richer web UI (graph visualization, quality scoring, 8 tabs); more permissive license (Apache 2.0); multiple transport options (stdio, SSE, HTTP)
• basic-memory — memories are human-readable Markdown files you can edit, version-control, and read without any server; wiki-style entity links form a local knowledge graph; ~20 MCP tools

When to pick Memlord:

• You want zero-config local embeddings — ONNX model ships with the server, no Ollama or external API needed
• You run a multi-user team server with proper OAuth 2.1 auth and invite-based workspaces
• You want a production-grade database (PostgreSQL) that scales beyond a single machine's SQLite
• You manage memories explicitly — store exactly what matters, typed and tagged, not everything the LLM decides to extract
• You want a self-hosted Web UI with full CRUD and JSON export, without a cloud subscription

🔍 How It Works

Each search request runs BM25 and vector KNN in parallel, then merges results via Reciprocal Rank Fusion:

flowchart TD
    Q([query]) --> BM25["BM25\nsearch_vector @@ websearch_to_tsquery"]
    Q --> EMB["ONNX embed\nall-MiniLM-L6-v2 · 384d · local"]
    EMB --> KNN["KNN\nembedding <=> query_vector\ncosine distance"]
    BM25 --> RRF["RRF fusion\nscore = 1/(k+rank_bm25) + 1/(k+rank_vec)\nk=60"]
    KNN --> RRF
    RRF --> R([top-N results])

🛠️ MCP Tools

Workspace management (create, invite, join, leave) is handled via the Web UI.

👨‍💻 Development

pyright src/           # type check
ruff format .          # format
pytest                 # run tests
alembic-autogen-check  # verify migrations are up to date

📄 License

Memlord is dual-licensed:

• AGPL-3.0 — free for open-source use. If you run a modified version as a network service, you must publish your source code.
• Commercial License — for proprietary or closed-source deployments. Contact sergey@memlord.com or dmitry@memlord.com to purchase.

🚀 Quickstart

🐳 Docker

cp .env.example .env
docker compose up

HTTP server (multi-user, Web UI, OAuth)

# Install dependencies
uv sync --dev

# Download ONNX model (~23 MB)
uv run python scripts/download_model.py

# Run migrations
alembic upgrade head

# Start the server
memlord

Open http://localhost:8000 for the Web UI. The MCP endpoint is at /mcp.

⚙️ Configuration

All settings use the MEMLORD_ prefix. See .env.example for the full list.

Set MEMLORD_BASE_URL to your public URL and change MEMLORD_OAUTH_JWT_SECRET before deploying.

💻 System Requirements

• Python 3.12
• PostgreSQL ≥ 15 with pgvector extension
• uv — Python package manager