Enterprise Self-Hosted
MemDB runs entirely on your infrastructure. Local ONNX embeddings — no OpenAI, no VoyageAI, no external API calls. Your conversations, your vectors, your knowledge graph — all behind your firewall.
Why Self-Host
Cloud memory APIs send every conversation to third-party servers. MemDB doesn't. Every component — embeddings, search, storage — runs on machines you control.
ONNX embeddings run locally. LLM extractor and reranker work with any OpenAI-compatible API — Ollama, vLLM, or your own GPU cluster. No vendor lock-in, no token metering.
No telemetry, no phone-home, no usage tracking. Audit the source code yourself — it's open source. SOC 2 and HIPAA compliance starts with not sending data out.
Bearer token auth with SHA-256 hashed master keys. Deploy behind VPN, in a private subnet, or on bare metal. MemDB is a Go binary — runs anywhere Linux runs.
Vectors (pgvector), graph (Apache AGE), and fulltext (tsvector) — all in one database. One backup strategy, one migration path, one ops team. No Pinecone, no Neo4j, no managed services.
Comparison
| Feature | MemDB | Mem0 | Zep | Letta |
|---|---|---|---|---|
| Self-hosted (free) | ✓ | Enterprise only | ✓ | ✓ |
| Local embeddings (no API) | ✓ ONNX | ✗ Requires OpenAI | ✗ Requires OpenAI | ✗ Requires OpenAI |
| Air-gapped / offline | ✓ | ✗ | ✗ | ✗ |
| MCP server built-in | ✓ stdio + HTTP | ✗ | ✗ | ✗ |
| Anthropic memory_20250818 adapter | ✓ Drop-in (v0.1.0) | ✗ | ✗ | ✗ |
| Self-hostable docker-compose | ✓ Single file | ✗ Cloud only (free tier) | ✓ OSS edition | ✓ |
| LoCoMo LLM Judge | 72.50% | 66.88% | 75.14% | — |
| License | Apache-2.0 | Apache-2.0 (cloud-gated) | Apache-2.0 + Commercial | Apache-2.0 |
| Single database (no graph DB) | ✓ PostgreSQL | Graph + Vector + KV | Neo4j + Postgres | ✓ Postgres |
| Hybrid search (vector + fulltext) | ✓ | Vector only | ✓ | Vector only |
| LLM reranker | ✓ Optional | ✗ | ✗ | ✗ |
| Temporal decay | ✓ Configurable | ✗ | ✓ | ✗ |
| Language | Go | Python | Go (cloud) / Python | Python |
| Cached search latency | 1.7ms | ~500ms | <200ms p95 | — |
Deployment
PostgreSQL, Qdrant, Redis, MemDB API, and MCP server — all containerized. No managed services, no cloud accounts, no API keys.
git clone https://github.com/anatolykoptev/memdb.git
cd MemDB
# Start everything (PostgreSQL + Qdrant + Redis + MemDB)
docker compose up -d
# Point LLM at local Ollama (or any OpenAI-compatible API)
# No cloud LLM required — use your own models
echo 'LLM_API_BASE=http://localhost:11434/v1' >> .env
echo 'LLM_MODEL=llama3' >> .env
# Verify
curl http://localhost:8080/health
# {"status":"healthy","service":"memdb-go"}
# Connect MCP to Claude Code
claude mcp add memdb -- ./memdb-mcp --stdio
# Embeddings: local ONNX. LLM: local Ollama. Data: local PostgreSQL.
# Nothing leaves your network. Self-hosted is free under Apache-2.0, forever. For enterprise support, custom SLAs, onboarding, or migration help — email consulting@memdb.ai.