Serein MCP
Serein keeps an AI's long-term memory on your own server and serves it to chats through MCP tools and an OpenAI-compatible gateway
- 231 ★ on the repository
- Last commit
Broad access
Why this level
- Stores personal conversations and diary entries, which cross the network in plain text when reached over unencrypted HTTP
- The gateway holds model keys and a shared access key, and MCP tools can edit and delete memory records
Install
- Works in
Manual install
bash scripts/one_click.shLinux and Termux, from the unpacked release. On Windows run `powershell -ExecutionPolicy Bypass -File .\scripts\one_click.ps1`. Needs Python 3.9 and Docker, or Python 3.11 and Node 22.12.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Stores personal conversations and diary entries, which cross the network in plain text when reached over unencrypted HTTP
- The gateway holds model keys and a shared access key, and MCP tools can edit and delete memory records
This is third-party code. Review the repository files before installing.
What it does
Serein is a personal memory service for AI chats that you deploy on your own machine. It provides a web interface, an OpenAI-compatible API gateway and an MCP server, so different chat windows read and write one shared memory. The public version starts from an empty database with no private memories or model keys. Memory splits into scenes and events. The main-window model writes a scene itself by calling `write_scene` when it judges a fragment worth keeping. Events are assembled by an auto-summary task from accepted messages, and the original messages stay archived for verification. Search blends vector and text matching, and a model can search on its own with `recall_memory` and read a full record with `read_memory`. There are also separate preference notes, a diary, narrative volumes (arcs), a candidate-memory queue and context handoff to a new window with the `/resume` command or the matching MCP tool. Some tools appear only after the feature is switched on in settings, and the list depends on the instance's permissions: a read-only instance keeps four basic tools. Data lives in SQLite with indexes stored separately, so the whole install directory has to be copied.
Who it is for. Owners of their own server who hold long conversations with a model and want the memory stored at home instead of at a vendor, connecting several chat clients to it.
Good fit when
- One conversation with a model sprawled over many windows and the memory has to carry over
- An MCP-capable client should search memories on its own through `recall_memory`
- Old chat history sits in an export and has to be loaded into memory and checked against the original messages
- A new window should pick up the previous conversation: `/resume` and the `resume` tool exist for that
- The model's memory has to live on your server together with backups
Not a fit when
- You need memory for code and repositories: Serein is built for personal conversations, not projects
- You cannot run your own server and maintain backups
- You need an English interface: the docs and interface are in Chinese
Example request
Use recall_memory to recall what we agreed on last week about the trip, then write the outcome as a scene with write_sceneMCP
- Transport
- http
- Authentication
- OAuth
Limitations
- The documentation and interface are in Chinese, and there is no English README.
- Auto-summary is not fully accurate: events can lose details or confuse references, so facts should be checked against the bound original messages.
- A client connected only through MCP does not import history automatically and has to call the tools itself.
- Plain HTTP on a public IP leaves the password and memories unencrypted, so public access needs HTTPS or a tunnel.
How to disable. Stop the service from the `se` menu in the install directory and delete that directory together with the backups. In your clients remove the MCP server and the gateway key.
FAQ
How do I connect Serein to an MCP client?
Set the address `https://your-domain/serein/mcp`, Streamable HTTP transport and OAuth sign-in: the gateway key is entered on the authorization page. A client without OAuth but with custom headers uses `Authorization: Bearer <gateway key>`.
What is the difference between a scene and an event?
The main model writes a scene on its own decision, while the auto-summary task assembles an event from accepted messages. Both take part in search, but the bound original messages are the facts about the conversation.
Which MCP tools are always available?
Four basic ones: `recall_memory`, `read_memory`, `find_arc` and `read_arc_materials`. The rest depend on the instance's permissions, enabled features and the tool allowlist.
What if I need to migrate from an older install?
The import page accepts chat history, Operit backups and old Ombre archives up to 64 MiB. A preview comes before writing, then the import runs in the background with pause and resume.
Does Serein need HTTPS?
OAuth sign-in for MCP needs a domain with HTTPS (or localhost). Over HTTP on a public IP only a static key in a header works, and the authors advise HTTPS or an encrypted tunnel for public access.
Discussion
No questions or reviews yet.
Related
MCP serversA long-term memory MCP server for AI agents: facts fade along a forgetting curve and related records are compressed into a summary
A self-hosted MCP memory server that keeps preferences and decisions across conversations in Claude Code, Cursor and other clients
A self-hosted MCP server for long-term agent memory: structured records, versioned rollback and a visual dashboard
iai-memory
iai-memory (Personal Memory Engine)
A local long-term memory MCP server for agents: keeps conversations verbatim and returns the right context at the start of a session
Membase keeps one shared memory for AI agents that Claude Code, Cursor and Codex read through MCP, an SDK or the membase command
ai-memory retains session knowledge in a Markdown wiki, exposes MCP retrieval and passes context between agents and machines.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Serein (this card) | no | OAuth | 231 | |
| Basic Memory | no | not required | 4,495 | notes in local Markdown files, with no gateway and no scene and event split |
| agentmemory | yes | not required | 30,160 | memory for coding agents, written by hooks and tied to projects |
README in short
The README opens with a description of personal memory for AI and a feature table: writing, recalling, preferences, arcs, data management, window switching. It then explains storage, the scene and event distinction, search and hooking up chat hosts. The page ends with a script install, a client connection table, the list of MCP tools and a link to the research paper on event memory.