MemCastle 发布本地优先的 AI 编码智能体共享记忆服务器

Local-first, always-on memory server for AI coding agents over MCP/HTTP

Running several AI coding agents (OpenCode, Claude Code, Cursor, ...) side by side usually means each one gets its own, disconnected memory — or none at all. MemCastle is a single daemon per palace (the memory store that one or more of your projects share) that all of them talk to over MCP or HTTP, so a mining run, a search, or a saved decision from one agent is immediately visible to every other agent and to the CLI. Everything lives in one SurrealDB store, embedded by default, so there is no external service to run.

flowchart LR
    A[OpenCode] & B[Claude Code] & C[Other MCP clients] -->|MCP| D
    CLI[memcastle CLI] -->|REST| D
    D[memcastle daemon] --> E[(Palace: SurrealDB)]

Installation

brew install noirbizarre/homebrew-tap/memcastle   # macOS
paru -S memcastle-bin                             # Arch Linux (AUR)

Or download a binary for your platform from the latest release, or build from source with cargo install --path .. See the installation guide for details and platform notes.

Quickstart

Start the daemon in one terminal. It serves 127.0.0.1:8420 and keeps its data in ~/.local/share/memcastle/default:

memcastle serve

Then, from another terminal:

memcastle status                       # is it up, where, and is its datastore healthy?
memcastle mine ./project               # submits a durable, resumable job — doesn't block
memcastle job list                     # follow it
memcastle search "why did we switch to GraphQL?"

Connect an MCP client to http://127.0.0.1:8420/mcp:

claude mcp add --transport http memcastle http://127.0.0.1:8420/mcp   # Claude Code

For OpenCode, add a remote server to opencode.json. The MCP client guide has the exact configuration for both. Stop and restart the daemon whenever you like: the palace is on disk, and it is all still there.

Documentation

The full documentation is at https://memcastle.github.io/.

To...	Read
Install and try it	Installation, Quickstart
Use it from an agent	Connect an MCP client, Agent integrations (`memcastle integration install`), Agent skills, MCP tools and REST API
Run it day to day	Running the daemon, Authentication, Memory modes, Project configuration, Mining sources, Deduplication, Database access, Troubleshooting
Configure it	Configuration (XDG paths on Linux and macOS, config file, environment, flags), CLI reference
Know where data lives	Storage and data, Migrations and upgrades
Understand or change it	Architecture, Development, Architecture Decisions, Writing and publishing sources, Integration contract

Contributing

See CONTRIBUTING.md.

License

MIT — see LICENSE.