Ori Mnemos MCP
Persistent agent memory in a folder of linked markdown notes, so Claude Code remembers a project across sessions
- 334 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- The agent reads and writes files in the vault, and Claude Code hooks run at every session start and end
- Conversation contents land in memory, so secrets mentioned in chats can end up in notes
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/ori-mnemosDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The agent reads and writes files in the vault, and Claude Code hooks run at every session start and end
- Conversation contents land in memory, so secrets mentioned in chats can end up in notes
This is third-party code. Review the repository files before installing.
What it does
Ori Mnemos keeps an agent's memory in a vault of plain markdown files. Notes are linked with wiki-links, and those links act as the edges of a knowledge graph. An SQLite index with vectors and graph metrics is built on top. The files stay the source of truth, and the index can be rebuilt with the `ori index build` command. Search reranks results by several signals: vector similarity, links between notes and recency. It also weighs which notes helped answer earlier questions. Recursive graph traversal with sub-question decomposition runs through the `ori explore` command. Embeddings come from the local all-MiniLM-L6-v2 model inside the process. The `ori bridge` command connects the vault to an agent. For Claude Code it writes hooks, the MCP server and CLAUDE.md, for Hermes Agent it installs a native plugin, and for OpenCode a JavaScript plugin and an AGENTS.md entry. In these three cases the agent gets a short briefing from memory at session start and saves insights at the end. The vault can live in git and move between machines. The server also runs on a VPS with no graphical interface.
Who it is for. For people running long projects in Claude Code, Hermes Agent or OpenCode who want the agent to keep decisions and context in human-readable files.
Good fit when
- The agent forgets project agreements between sessions and you tire of repeating them at the start of every chat
- You want memory you can open in an editor, edit by hand and keep in git
- You want one shared memory for several clients, for example Claude Code and Cursor
- You need the agent to forget or replace an outdated fact on command, without rebuilding the whole vault
Not a fit when
- You want ready cloud memory with a web UI and accounts: Ori works on your files and hosts nothing for you
- You need real-time shared memory for a team: the vault lives on disk and syncing is left to git
- You have few notes and a short project: a rules file in the repository is enough
- Memory must reliably erase secrets: hard deletion of a fact has to be checked by hand
Example request
What did we decide about caching in this project? Search memory and show the linked notes.MCP
- Transport
- stdio
- Authentication
- not required
Environment variables
ORI_VAULTrequired- Path to the vault folder the server works with
Limitations
Full session automation, meaning a briefing at the start and saving at the end, exists only for Claude Code, Hermes Agent and OpenCode. Cursor, Codex and other MCP clients get the tools without that automation.
For Hermes Agent with Docker or SSH remote terminals, the ori command must be installed inside the environment where the agent runs, and the vault must be reachable from there. Connecting a language model is optional: without it every operation works on heuristics.
The comparisons with other memory systems were prepared by the author, who also lists their limits. Those numbers are worth rechecking on your own data.
How to disable. The `ori bridge` command with the `--uninstall` flag removes what Ori owns from supported clients: hooks, plugins and the MCP entry. For the generic variant Ori only prints removal instructions, because it does not touch a config it does not own. The notes stay: the vault is an ordinary folder you can remove by hand.
FAQ
How do I connect Ori Mnemos to Claude Code?
Install the package with `npm install -g ori-memory`, create the vault with `ori init my-agent` and run `ori bridge claude-code --vault ~/brain`. Running it again with the same agent updates the vault path or activation mode in place.
What does Ori Mnemos store and where?
Notes sit in markdown files in the vault folder, and wiki-links define the connections. An SQLite index is built next to them and can be rebuilt from the files. The scores and retrieval history that ranking learns from live only in the index, so export them before deleting it.
Does Ori Mnemos need an API key?
Not for basic work: embeddings are computed locally, and search and graph traversal run without a network. An Anthropic, OpenAI-compatible or local language model can be connected in the llm section of ori.config.yaml. It improves classification and link suggestions.
How does the global install scope differ from project?
With global, one vault applies across the machine; with project it stays inside one repository. A project install overrides the global one, and an explicit `--vault` overrides an inferred path. The auto activation mode runs the briefing at session start, while manual only leaves the tools available.
How do I ask the agent to forget a fact?
There are operations for superseding a fact, soft eviction on request and hard deletion. Hard deletion is the weakest of the three, so for secrets check the result by hand.
Discussion
No questions or reviews yet.
Related
MCP serversLong-term agent memory in local Markdown files with a link graph, semantic search and MCP access
One local binary that gives Claude Code, Codex and Cursor shared memory across sessions, with a knowledge graph on disk and no API keys
Local persistent memory for coding agents: hooks record every session, and Claude Code, Codex or Cursor recall past decisions through 54 MCP tools
A long-term memory MCP server for agents: one Rust binary that runs offline and never calls a language model
Memory for coding agents in plain Markdown: the agent keeps specs, decisions and lessons itself, and each new session starts from them. No vector database
Long-term memory for agents: it distills knowledge from session history and brings it back in new tasks, one store shared across Claude Code, Codex, Cursor and
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Ori Mnemos (this entry) | no | not required | 334 | |
| Hindsight | yes | API key | 55,871 | also stores agent experience between sessions, while Ori keeps memory in markdown files with no separate system |
| M-flow | no | API key | 4,510 | a knowledge-graph memory MCP server, while Ori builds its graph from the wiki-links of ordinary notes |
README in short
Most of the documentation is the author's comparison with Mem0, LangMem and other systems on four benchmark sets, with caveats about weak spots. The rest is a reference for MCP tools, CLI commands and ori.config.yaml parameters.