supermem v0.1.0 — Initial Release
Persistent AI memory without RAG — four-tier retrieval that uses an LLM agent only as a last resort, backed by SQLite FTS5, an embedded graph database, and your local markdown vault.
What is supermem?
supermem is an MCP (Model Context Protocol) server that gives AI assistants — Claude Desktop, LM Studio, ChatGPT — persistent, structured memory backed entirely on your local machine. Most queries resolve in milliseconds via full-text search; the LLM agent is a last resort, not the default path.
Installation
pip install recallQuick Start (Personal, No GPU)
export RECALL_VAULT_PATH=~/notes
export RECALL_LLM_PROVIDER=openrouter
export OPENROUTER_API_KEY=your_key_here
recall serveAdd to Claude Desktop mcp.json:
{
"mcpServers": {
"recall": {
"command": "recall",
"args": ["serve"]
}
}
}Quick Start (Docker)
git clone https://github.com/lamenting-hawthorn/supermem
cp .env.example .env
# Edit .env with your vault path and API keys
docker compose up recall-mcpArchitecture: Four-Tier Retrieval
Query
│
├─ Tier 1: SQLite FTS5 full-text search ~1ms always available
├─ Tier 2: Kuzu embedded graph expansion ~5ms optional
├─ Tier 3: ChromaDB vector similarity ~50ms optional
└─ Tier 4: LLM agent fallback ~5-30s always available
Tiers 1–3 never call an LLM. If tier 1 returns enough results, the rest are skipped entirely.
Features
- MCP server — stdio and HTTP transports; works with Claude Desktop, LM Studio, ChatGPT
- Hybrid storage — SQLite FTS5, Kuzu graph database, ChromaDB vector store
- Ambient memory injection — Claude Code hook system surfaces relevant context automatically
- TTL-based expiry — configure retention via
RECALL_OBS_TTL_DAYS(default: 90 days) - Parallel retrieval — all tiers run concurrently for sub-100ms queries
- Memory import connectors — ChatGPT, Notion, Nuclino, GitHub, Google Docs
- Worker HTTP API — search, indexing, health checks on
:37777 recallCLI —serve,chat,backup,restore,connectcommands- Docker support — single-command setup with
docker compose
Requirements
- Python 3.11
- Optional:
kuzufor graph tier,chromadbfor vector tier, Docker for containerized setup