Deploy and Host MemPalace Remote MCP Hub on Railway

MemPalace is an open source AI memory system. Agents file what matters into a "palace" of wings, rooms and drawers, store it verbatim, and recall it later with semantic search and a temporal knowledge graph. This template runs it as a remote MCP server so every agent and teammate shares one memory over HTTPS.

About Hosting MemPalace Remote MCP Hub

Hosting MemPalace is a single container. This template runs the official ghcr.io/mempalace/mempalace:3.10.0 image with upstream's own network server command, mempalace serve --host 0.0.0.0 --port 8765, and a Railway volume at /data that holds the ChromaDB palace, the knowledge graph, the write ahead log and the cached embedding model. A bearer token is generated at deploy and required on every route except /healthz; without it the server refuses to start, so the memory store is never open to the internet. PORT is pinned to 8765 so the Railway healthcheck hits /healthz, RAILWAY_RUN_UID=0 lets the non root image write the root owned volume, and the idle watchdog that upstream uses to reap desktop sessions is turned off so the server stays up.

Common Use Cases

Shared memory for coding agents: Claude Code, Cursor, Codex and other MCP clients read and write the same palace
Long term memory across sessions: decisions, preferences and project facts survive restarts and machine changes
Team knowledge base: one hub for a team, with verbatim storage instead of lossy summaries
Agent coordination: the built in logstream lets agents hand tasks and patches to each other
Memory backend for agents on Railway: OpenClaw, Hermes or your own agent in the same project

Dependencies for MemPalace Remote MCP Hub Hosting

Railway volume at /data (included)
No API keys: embeddings run locally on CPU with the minilm ONNX model, downloaded once (about 80 MB) on the first write
No database service: the default ChromaDB backend lives on the volume. Qdrant is an optional upgrade

Deployment Dependencies

Implementation Details

The start command is docker-entrypoint.sh serve --host 0.0.0.0 --port 8765, the same command as upstream's deploy/docker-compose.server.yml. Key variables:

PORT=8765
MEMPALACE_MCP_HTTP_TOKEN=${{secret(32)}}
MEMPALACE_MCP_IDLE_HOURS=0
RAILWAY_RUN_UID=0
MEMPALACE_CONFIG_DIR=/data/.mempalace
MEMPALACE_EMBEDDING_MODEL=minilm

First steps after deploy

Wait for https:///healthz to return ok.
Copy MEMPALACE_MCP_HTTP_TOKEN from the service variables.
Connect Claude Code: claude mcp add --transport http mempalace https:///mcp --header "Authorization: Bearer ". Other clients use the same URL and header.
Ask your agent to remember something. The first write downloads the embedding model, so give it up to a minute; later calls are fast.

Keep one replica: the palace has a single writer. The token grants full read and write access; set MEMPALACE_MCP_READ_ONLY=1 for a recall only server. Choose the embedding model before the first write, since switching later means rebuilding the index. About 300 MB of RAM is used once the model is loaded, so the Hobby plan is recommended.

Why Deploy MemPalace Remote MCP Hub on Railway?

Railway is a singular platform to deploy your infrastructure stack. Railway will host your infrastructure so you don't have to deal with configuration, while allowing you to vertically and horizontally scale it.

By deploying MemPalace Remote MCP Hub on Railway, you get a pinned, health checked, token protected memory server with persistent storage, managed TLS and private networking to the rest of your Railway project.