Deploy AutoMem | (Just Updated) Agent Memory That Actually Embeds

Deploy and Host AutoMem on Railway

AutoMem is a self-hosted memory service for AI coding agents — Claude Code, Codex CLI, Cursor, Cline. It stores what your agents learn as a knowledge graph (FalkorDB) plus a vector index (Qdrant), and serves recall over a REST API and an MCP bridge, so an agent can remember decisions, preferences and project facts across sessions and across tools.

About Hosting AutoMem

This template runs four services: the AutoMem API, the FalkorDB graph store on a volume, Qdrant on a volume, and the MCP bridge your agent connects to. Two things about a hosted AutoMem are easy to get wrong, and this packaging fixes both.

Embeddings. With no hosted embedding key, AutoMem falls back to a local model — BAAI/bge-large-en-v1.5, about 1.2 GB — which the stock image downloads at container start into a path that is not on a volume, so the download repeats on every deploy. If that download fails, AutoMem's provider chain falls through one more step to a placeholder provider that returns random.Random(sha256(text)) vectors: stores succeed, recalls succeed, /health reports healthy, and the results are noise. This image bakes the model in at build time (outside any mount, so the volume cannot hide it) and pins the provider to the local model, which raises rather than degrading silently. Boot measured 3.9 seconds to "ready" instead of ~35 seconds of download.

Graph authentication. The FalkorDB image does not read FALKOR_PASSWORD; only REDIS_ARGS reaches redis-server. A template that generates a password and passes it to the API still boots a graph database that logs "Redis does not require authentication" and answers any client that can reach it. This image appends --requirepass itself and refuses to boot without a password. It also disables the bundled Next.js graph browser, which upstream starts by default.

Everything is pinned — AutoMem 0.16.1, FalkorDB v4.20.4, Qdrant v1.19.0 — and the deploy form asks for nothing: tokens and the graph password are generated per deployment.

Why Deploy AutoMem on Railway?

Railway gives each service private networking, so only the API and the MCP bridge get public URLs while the graph and vector stores stay inside the project. Volumes keep the graph and the vector index across redeploys, and usage-based billing suits a service that is idle between agent calls.

Common Use Cases

• Persistent memory for Claude Code, Codex CLI, Cursor or Cline over MCP
• A shared memory backend for several agents or several machines
• Semantic recall over project decisions, preferences and past sessions

Dependencies for AutoMem

• FalkorDB (graph store, included)
• Qdrant (vector store, included)
• Optional: a Voyage or OpenAI API key for hosted embeddings instead of the bundled local model

Deployment Dependencies

• AutoMem
• FalkorDB
• Qdrant