MemoryMesh

MemoryMesh is a knowledge graph server designed for AI models, with a focus on text-based RPGs and interactive storytelling. It helps AI maintain consistent, structured memory across conversations, enabling richer and more dynamic interactions.

The project is based on the Knowledge Graph Memory Server from the MCP servers repository and retains its core functionality.

IMPORTANT

v0.4.0 Update: The internals were rewritten to be much smaller and more reliable. Tool names, schemas and the memory file format are unchanged, so your existing setup and data keep working. Notable fixes: rebuilding or running npm install no longer erases dist/data/memory.json, and a failed tool call no longer leaves half-saved data behind. See the changelog for details.

Schemas live in dist/data/schemas and memory in dist/data/memory.json by default. Both locations can be changed with environment variables.

Quick Links

• Installation
• Example
• SchemaManager Guide Discussion
• MemoryViewer Guide Discussion

Overview

MemoryMesh is a local knowledge graph server that empowers you to build and manage structured information for AI models. While particularly well-suited for text-based RPGs, its adaptable design makes it useful for various applications, including social network simulations, organizational planning, or any scenario involving structured data.

Key Features

• Dynamic Schema-Based Tools: Define your data structure with schemas, and MemoryMesh automatically generates tools for adding, updating, and deleting data.
• Intuitive Schema Design: Create schemas that guide the AI in generating and connecting nodes, using required fields, enumerated types, and relationship definitions.
• Metadata for AI Guidance: Use metadata to provide context and structure, helping the AI understand the meaning and relationships within your data.
• Relationship Handling: Define relationships within your schemas to encourage the AI to create connections (edges) between related data points (nodes).
• Informative Feedback: Provides error feedback to the AI, enabling it to learn from mistakes and improve its interactions with the knowledge graph.
• Event Support: An event system tracks operations, providing insights into how the knowledge graph is being modified.

Nodes

Nodes represent entities or concepts within the knowledge graph. Each node has:

• name: A unique identifier.
• nodeType: The type of the node (e.g., npc, artifact, location), defined by your schemas.
• metadata: An array of strings providing descriptive details about the node.

Example Node:

    {
      "name": "Aragorn",
      "nodeType": "player_character",
      "metadata": [
        "Race: Human",
        "Class: Ranger",
        "Skills: Tracking, Swordsmanship",
        "Affiliation: Fellowship of the Ring"
      ]
    }

Edges

Edges represent relationships between nodes. Each edge has:

• from: The name of the source node.
• to: The name of the target node.
• edgeType: The type of relationship (e.g., owns, located_in).
• weight: (Optional) A number between 0 and 1 representing the strength or confidence of the relationship. Treated as 1 when not set.

{
  "from": "Aragorn",
  "to": "Andúril",
  "edgeType": "owns"
}

Schemas

Schemas are the heart of MemoryMesh. They define the structure of your data and drive the automatic generation of tools.

Schema File Location

Place your schema files (.schema.json) in the dist/data/schemas directory of your built MemoryMesh project. MemoryMesh will automatically detect and process these files on startup.

Schema Structure

File name: any name ending in .schema.json, for example npc.schema.json. The tool names come from the name field inside the file, not from the file name.

• name - Identifier for the schema and node type within the memory. IMPORTANT: The schema’s name must start with add_ to be recognized.
• description - Used as the description for the add_<name> tool, providing context for the AI. (The delete and update tools have a generic description)
• properties - Each property includes its type, description, and additional constraints.
  • property
    • type - Supported values are string or array.
    • description - Helps guide the AI on the entity’s purpose.
    • required - Boolean. If true, the AI is forced to provide this property when creating a node.
    • enum - An array of strings. If present, the AI must choose one of the given options.
    • relationship - Defines a connection to another node. The AI provides one or more node names, and MemoryMesh creates an edge to each of them. If a property is required and has a relationship, the AI will always create both the node and the corresponding edge.
      • edgeType - Type of the relationship to be created.
      • description - Helps guide the AI on the relationship’s purpose.
      • The target nodes must already exist. If one doesn't, the whole call fails with an explanation and nothing is saved, so the AI can create the missing node and try again.
• additionalProperties - Boolean. If true (or left out), allows the AI to add extra attributes beyond those defined in properties. If false, extra attributes are rejected.

If a schema file is invalid, MemoryMesh skips it, logs the reason, and still starts with the other schemas.

Example Schema (add_npc.schema.json):

{
  "name": "add_npc",
  "description": "Schema for adding an NPC to the memory" ,
  "properties": {
    "name": {
      "type": "string",
      "description": "A unique identifier for the NPC",
      "required": true
    },
    "race": {
      "type": "string",
      "description": "The species or race of the NPC",
      "required": true,
      "enum": [
        "Human",
        "Elf",
        "Dwarf",
        "Orc",
        "Goblin"
      ]
    },
    "currentLocation": {
      "type": "string",
      "description": "The current location of the NPC",
      "required": true,
      "relationship": {
        "edgeType": "located_in",
        "description": "The current location of the NPC"
      }
    }
  },
  "additionalProperties": true
}

Based on this schema, MemoryMesh automatically creates:

• add_npc: To add new NPC nodes.
• update_npc: To modify existing NPC nodes. Each field you pass replaces that field's entry in the node's metadata; other entries are kept. Passing metadata replaces the whole list.
• delete_npc: To remove NPC nodes (and their edges).

MemoryMesh includes 11 pre-built schemas designed for text-based RPGs, providing a ready-to-use foundation for game development.

SchemaManager Tool

MemoryMesh includes a SchemaManager tool to simplify schema creation and editing. It provides a visual interface, making it easy to define your data structures without writing JSON directly.

Dynamic Tools

MemoryMesh simplifies interaction with your knowledge graph through dynamic tools. These tools are not manually coded but are automatically generated directly from your schema definitions. This means that when you define the structure of your data using schemas, MemoryMesh intelligently creates a set of tools tailored to work with that specific data structure.

Think of it like this: You provide a blueprint (the schema), and MemoryMesh automatically constructs the necessary tools to build, modify, and remove elements based on that blueprint.

How does it work behind the scenes?

MemoryMesh has an intelligent system that reads your schema definitions. It analyzes the structure you've defined, including the properties of your entities and their relationships. Based on this analysis, it automatically creates a set of tools for each entity type:

• add_<entity>: A tool for creating new instances of an entity.
• update_<entity>: A tool for modifying existing entities.
• delete_<entity>: A tool for removing entities.

These tools are then made available through a central hub within MemoryMesh, ensuring they can be easily accessed and used by any connected client or AI.

In essence, MemoryMesh's dynamic tool system provides a powerful and efficient way to manage your knowledge graph, freeing you to focus on the content and logic of your application rather than the underlying mechanics of data manipulation.

Memory file

By default, data is stored in dist/data/memory.json (one JSON object per line). You can move it anywhere with the MEMORYMESH_MEMORY_FILE environment variable, which is a good idea if you want it outside the build folder.

Memory Viewer

The Memory Viewer is a separate tool designed to help you visualize and inspect the contents of the knowledge graph managed by MemoryMesh. It provides a user-friendly interface for exploring nodes, edges, and their properties.

Key Features:

• Graph Visualization: View the knowledge graph as an interactive node-link diagram.
• Node Inspection: Select nodes to see their nodeType, metadata, and connected edges.
• Edge Exploration: Examine relationships between nodes, including edgeType and direction.
• Search and Filtering: Quickly find specific nodes or filter them by type.
• Table View: Allows you to easily find and inspect specific nodes and edges, or all of them at once.
• Raw JSON View: Allows you to view the raw JSON data from the memory file.
• Stats Panel: Provides key metrics and information about the knowledge graph: total nodes, total edges, node types, and edge types.
• Search and Filter: Allows you to filter by node type or edge type and filter whether to show nodes, edges, or both.

Accessing the Memory Viewer

The Memory Viewer is a standalone web application. Memory Viewer discussion

Using the Memory Viewer

• Select Memory File: In the Memory Viewer, click the "Select Memory File" button.
• Choose File: Navigate to your MemoryMesh project directory and select the memory.json file (located in dist/data/memory.json by default).
• Explore: The Memory Viewer will load and display the contents of your knowledge graph.

Memory Flow

Prompt

For optimal results, use Claude's "Projects" feature with custom instructions. Here's an example of a prompt you can start with:

You are a helpful AI assistant managing a knowledge graph for a text-based RPG. You have access to the following tools: add_npc, update_npc, delete_npc, add_location, update_location, delete_location, and other tools for managing the game world.

When the user provides input, first process it using your available tools to update the knowledge graph. Then, respond in a way that is appropriate for a text-based RPG.

You can also instruct the AI to perform specific actions directly in the chat.

Experiment with different prompts to find what works best for your use case!

Example

1. A simple example with custom instructions.
2. An example for the sake of example, with visualization (NOT part of the functionality)

Add a couple of cities, some npcs, couple locations around the city to explore, hide an artifact or two somewhere

Installation

Prerequisites

• Node.js: Version 18 or higher (a current LTS version is recommended). You can download it from nodejs.org.
• npm: Usually included with Node.js.
• Claude for Desktop: Make sure you have the latest version installed from claude.ai/download.

Installation Steps

1. Clone the Repository:

   git clone https://github.com/CheMiguel23/memorymesh.git
   cd memorymesh
   

2. Install Dependencies:

   npm install
   

3. Build the Project:

   npm run build
   

   This command compiles the TypeScript code into JavaScript in the dist directory and copies the example schemas to dist/data/schemas. Schemas that already exist there are not overwritten, so your edits survive rebuilds.

4. Verify the Build (Optional):

   • Check that dist/index.js exists and that dist/data/schemas contains .schema.json files.

5. Configure Claude Desktop:

   Open your Claude Desktop configuration file:

   • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
   • Windows: %APPDATA%\Claude\claude_desktop_config.json
   • Add an entry for memorymesh to the mcpServers section. You can choose one of the following configuration options:

   "mcpServers": {
     "memorymesh": {
       "command": "node", 
       "args": ["/ABSOLUTE/PATH/TO/YOUR/PROJECT/memorymesh/dist/index.js"]
     }
   }
   

   • Replace /ABSOLUTE/PATH/TO/YOUR/PROJECT/ with the actual absolute path to your memorymesh project directory.
   • Example (macOS):

     "command": "node",
     "args": ["/Users/yourusername/Projects/memorymesh/dist/index.js"]
     

   • Example (Windows):

     "command": "node",
     "args": ["C:\\Projects\\memorymesh\\dist\\index.js"]
     

6. Restart Claude Desktop: Completely restart Claude Desktop for the changes to take effect.

Verify Installation

1. Start Claude Desktop.
2. Open a new chat.
3. Look for the MCP plugin icon in the top-right corner. If it's there, your configuration is likely correct.
4. Click the icon. You should see "memorymesh" in the list of connected servers.
5. Click the icon. If you see tools listed (e.g., add_npc, update_npc, etc.), your server is working and exposing tools correctly.

Updating

git pull
npm install

Then restart your MCP client. Since v0.4.0, updating keeps your memory file and any schemas you edited in dist/data/schemas. Backing up dist/data first is still a good habit.

Troubleshooting

• Server not appearing in Claude:

  • Double-check the paths in your claude_desktop_config.json. Make sure they are absolute paths and correct.
  • Verify that the dist directory exists and contains the compiled JavaScript files, including index.js.
  • Check the Claude Desktop logs for errors:
    • macOS: ~/Library/Logs/Claude/mcp-server-memorymesh.log (and mcp.log)
    • Windows: (Likely in a Logs folder under %AppData%\Claude)

• Tools not showing up:

  • Make sure your npm run build command completed without errors.
  • Verify that your schema files are in dist/data/schemas (or your MEMORYMESH_SCHEMAS_DIR), end in .schema.json, and have a name starting with add_.
  • Check the server log: on startup MemoryMesh prints which schemas it loaded, and the reason for any schema it skipped.

Advanced Configuration

Environment Variables

Set them in your MCP client configuration, for example in Claude Desktop:

"mcpServers": {
  "memorymesh": {
    "command": "node",
    "args": ["/ABSOLUTE/PATH/TO/memorymesh/dist/index.js"],
    "env": {
      "MEMORYMESH_MEMORY_FILE": "/ABSOLUTE/PATH/TO/my-campaign/memory.json",
      "MEMORYMESH_SCHEMAS_DIR": "/ABSOLUTE/PATH/TO/my-campaign/schemas"
    }
  }
}

This also makes it easy to keep separate memories (for example, one per campaign) by configuring several servers with different files.

Docker

docker build -t memorymesh .
docker run -i --rm -v memorymesh-data:/data memorymesh

Inside the container the memory file is /data/memory.json, so mount a volume there to keep it.

Limitations

1. Node Deletion: The AI may be hesitant to delete nodes from the knowledge graph. Encourage it through prompts if needed.

2. Conflicting Knowledge: MemoryMesh currently uses a "last write wins" approach for handling data. If conflicting information is provided about the same entity, the most recent update will overwrite previous values. Here are strategies for managing conflicting information:

   Current Approaches:

   • Use metadata to track sources: Add metadata entries like "Source: Character testimony" or "Source: Official records" to track where information came from.
   • Use temporal metadata: Include timestamps or narrative time markers (e.g., "As of Chapter 3") in metadata to track when information was valid.
   • Create separate nodes for perspectives: For subjective or disputed information, create separate nodes representing different viewpoints (e.g., rumor_about_villain vs truth_about_villain).
   • Use edge weights: Leverage the optional weight property on edges (0-1 range) to indicate confidence or reliability of relationships.

   Example - Tracking uncertain information:

   {
     "name": "VillainOrigin_Rumor",
     "nodeType": "information",
     "metadata": [
       "Source: Tavern gossip",
       "Reliability: Low",
       "Claims: Villain came from the northern mountains"
     ]
   }
   

   Future Considerations:
   For applications requiring sophisticated conflict resolution, consider implementing a custom layer that:

   • Maintains version history of node changes
   • Tracks provenance (source) of each piece of information
   • Implements confidence scores for assertions
   • Supports temporal validity periods for facts

Development

npm install     # also builds
npm test        # runs the test suite
npm run watch   # recompiles on change

The source is small:

Contribution

Contributions, feedback, and ideas are welcome!
This project is a personal exploration into integrating structured data with AI reasoning capabilities. Contributions, feedback, and ideas are welcome to push it further or inspire new projects.