MemoryMesh is an MCP server that gives AI models a structured, long-term memory in the form of a knowledge graph. You describe the kinds of things you want to remember in small schema files (people, projects, characters, locations, books...), and MemoryMesh turns each one into tools the AI uses to add, update and delete that information, linking related items as it goes.
It works for anything you can describe as things and the connections between them: a text-based RPG world, the people and teams at your company, a personal CRM, or your reading notes.
Based on the Knowledge Graph Memory Server from the MCP servers repository.
The memory is a graph of nodes and edges:
- A node is a thing:
{"name": "Mara", "nodeType": "npc", "metadata": ["role: Blacksmith", "status: injured"]} - An edge is a connection between two nodes:
{"from": "Mara", "to": "Riverton", "edgeType": "located_in"}. Edges can have an optionalweightfrom 0 to 1 (strength or confidence; treated as 1 when not set).
A schema describes one type of node. For a schema named add_npc, MemoryMesh creates three tools:
add_npccreates an NPC. Fields marked as relationships also create edges to the nodes they name.update_npcchanges the fields you pass and keeps everything else. Passingmetadatareplaces the whole list.delete_npcremoves the NPC and its edges.
Every call is all-or-nothing: if anything is wrong (a missing required field, a link to a node that doesn't exist), nothing is saved, and the AI gets a clear error it can act on.
Each schema is a JSON file ending in .schema.json. This is the npc example that ships with MemoryMesh:
{
"name": "add_npc",
"description": "Add a new character (NPC) to the knowledge graph",
"properties": {
"name": { "type": "string", "description": "The character's unique name", "required": true },
"description": { "type": "string", "description": "Who the character is", "required": true },
"status": { "type": "string", "description": "Current condition", "enum": ["alive", "injured", "dead", "missing"] },
"traits": { "type": "array", "description": "Personality traits" },
"currentLocation": {
"type": "string",
"description": "Where the character is now",
"relationship": { "edgeType": "located_in", "description": "Where the character is now (must already exist)" }
}
},
"additionalProperties": true
}| Key | Meaning |
|---|---|
name |
Must start with add_. The rest becomes the node type and tool suffix (npc → add_npc, update_npc, delete_npc). |
description |
Description of the add_ tool, shown to the AI. |
properties |
The fields. Each has a type (string or array) and a description, plus optionally: |
↳ required |
true if the AI must provide this field. Keep this to a minimum, or the AI will invent values it doesn't know. |
↳ enum |
A list of allowed values. |
↳ relationship |
Creates an edge (edgeType) to each node named in this field. Those nodes must already exist. |
additionalProperties |
true (default) saves fields not listed above as extra metadata; false rejects them. |
If a schema file is invalid, MemoryMesh skips it, logs the reason, and starts with the others.
The SchemaManager lets you create and edit schemas visually (guide).
MemoryMesh ships with two small examples, npc and location. Ready-made sets for different uses are shared in Discussions:
- People & organization: people, teams, projects, places and events (who reports to whom, who's responsible for what, who went where, birthdays)
- Personal CRM: contacts, companies, interactions and follow-ups
- Research & reading notes: sources, authors, concepts and notes
- RPG & storytelling: the full set of 11 schemas for text-based games
Feel free to share your own.
These are always available, whatever schemas you use:
| Tool | Purpose |
|---|---|
read_graph |
Return the whole graph |
search_nodes |
Find nodes by name, type or metadata, with their direct neighbours |
open_nodes |
Return specific nodes by name, with their direct neighbours |
add_nodes, update_nodes, delete_nodes |
Work with nodes of any type |
add_edges, update_edges, delete_edges |
Work with edges |
add_metadata, delete_metadata |
Add or remove individual metadata entries |
You need Node.js 18 or newer and an MCP client such as Claude Desktop.
-
Download and build MemoryMesh:
git clone https://github.com/CheMiguel23/MemoryMesh.git
cd MemoryMeshnpm install
npm installalso builds the project intodist/. -
Add MemoryMesh to Claude Desktop. Open Settings → Developer → Edit Config and add:
{ "mcpServers": { "memorymesh": { "command": "node", "args": ["/ABSOLUTE/PATH/TO/MemoryMesh/dist/index.js"] } } }On Windows, double every backslash:
"C:\\Projects\\MemoryMesh\\dist\\index.js". -
Fully quit and restart Claude Desktop. MemoryMesh and its tools (
add_npc,read_graph, ...) appear in the tools menu of the chat box.
git pullnpm installThen restart your MCP client. Updating keeps your memory file and any schemas you edited.
| Environment variable | What it sets | Default |
|---|---|---|
MEMORYMESH_MEMORY_FILE |
Path of the memory file | dist/data/memory.json |
MEMORYMESH_SCHEMAS_DIR |
Folder containing .schema.json files |
dist/data/schemas |
Set them in your Claude Desktop config:
"memorymesh": {
"command": "node",
"args": ["/ABSOLUTE/PATH/TO/MemoryMesh/dist/index.js"],
"env": {
"MEMORYMESH_MEMORY_FILE": "/ABSOLUTE/PATH/TO/my-project/memory.json",
"MEMORYMESH_SCHEMAS_DIR": "/ABSOLUTE/PATH/TO/my-project/schemas"
}
}Keeping your schemas and memory in a folder of your own keeps them separate from the MemoryMesh code. You can also run several MemoryMesh entries with different folders, for example one per campaign or project.
Docker: docker build -t memorymesh . then docker run -i --rm -v memorymesh-data:/data memorymesh. The memory file is /data/memory.json inside the container.
The memory file stores one JSON object per line. The MemoryViewer displays it as a searchable table with stats (guide): open it in Chrome or Edge, click Select Memory File, and pick your memory.json.
- Tell the AI what the memory is for, for example in a Claude Project's instructions: "Use the memorymesh tools to record every person, team and project I mention, and check the graph before answering questions about them."
- The AI can be hesitant to delete things. Ask explicitly when you want something removed.
- When information conflicts, the latest update wins. To keep both versions, record the source or date in metadata (
"source: tavern gossip","as of: chapter 3"), or store competing claims as separate nodes.
- MemoryMesh doesn't appear in Claude Desktop: check that the path in the config is absolute and points to
dist/index.js(with doubled backslashes on Windows), and thatnpm installfinished without errors. - Tools are missing: check that your schema files end in
.schema.jsonand that theirnamestarts withadd_. - Check the log, where MemoryMesh reports which schemas it loaded and why any were skipped:
- macOS:
~/Library/Logs/Claude/mcp-server-memorymesh.log - Windows: the
logsfolder inside%APPDATA%\Claude
- macOS:
npm install # also builds
npm test # runs the test suite
npm run watch # recompiles on change| File | Purpose |
|---|---|
src/index.ts |
Entry point: loads schemas and starts the server on stdio |
src/server.ts |
Lists tools and routes each tool call to its handler |
src/tools.ts |
Definitions of the built-in tools |
src/schemas.ts |
Loads schema files and implements the generated add_/update_/delete_ tools |
src/graph.ts |
The graph operations themselves, with input validation |
src/store.ts |
Reads and writes the memory file, one change at a time |
src/config.ts |
Paths and version |
src/types.ts |
Shared types |
Contributions, feedback, ideas and schema sets are welcome. Open an issue, start a discussion, or send a pull request.