KnowledgeSmith

KnowledgeSmith MCP Server enables Graphiti memory and document chunking by storing document chunks as episodes in a knowledge graph. It provides intelligent chunking, incremental sync, and a drop-in replacement for graphiti-memory APIs, so you can search memory nodes and facts and manage memory episodes with ease.

How to use

You run the MCP server locally or remotely and connect your Graphiti-enabled client to it. The server exposes memory graph capabilities and document chunking so you can add documents, query memory, and manage memory episodes. Use an MCP client that supports stdio or HTTP endpoints and point it at the server via the configured command, URL, or local process. You will typically initialize the server with the required environment variables (Neo4j connection, OpenAI API key, and a root directory for documents) and then use the provided tools to add documents and query the graph.

How to install

Prerequisites include a running Neo4j database and an OpenAI API key for embeddings. You also need a runtime environment to install and run the MCP server (see options below).

Install Neo4j (recommended via Docker):

# Run Neo4j with default password
docker run \
  -p 7474:7474 \
  -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/your-password \
  --name neo4j \
  neo4j:latest

Verify the database is reachable at http://localhost:7474 and that you can login with the user neo4j and the password you chose.

Get the OpenAI API key from your OpenAI account. You will use this for Graphiti embeddings.

Install the MCP server (two available options):

# Option 1: Install from source using uv
git clone https://github.com/yourusername/KnowledgeSmith.git
cd KnowledgeSmith
uv pip install -e .

# Option 2: Direct installation via uv
uv pip install rbt-mcp-server

Configure your client or Claude Desktop to run the MCP server with the required environment variables (see Quick Start for exact configurations).

Additional notes

The server supports 8 MCP tools for document and memory graph management, including adding documents, searching memory nodes and facts, retrieving episodes, and managing edges. Ensure your runtime environment provides the needed environment variables and that your client is prepared to handle stdio or HTTP connections as configured.

Security and maintenance

Keep your Neo4j password and OpenAI API key secure. Rotate credentials periodically and limit access to the MCP server endpoint. Regularly update dependencies to receive security fixes and performance improvements.

Test the server

Start the server after configuration and run a quick check to ensure the process is healthy and responsive. You should be able to connect from your MCP client and perform a simple document add or memory query.

# Example start command (if using a standard local run)
rbt-mcp-server

Available tools

add_document

Sync documents to the knowledge graph with automatic chunking by document structure; incremental updates only touch changed chunks.

search_memory_nodes

Query memory graph nodes such as entities, preferences, and procedures.

search_memory_facts

Query knowledge graph facts and relationships.

get_episodes

Retrieve recent memory episodes from the graph.

delete_episode

Remove a specific memory episode from the graph.

get_entity_edge

Fetch a specific entity relationship edge by UUID.

delete_entity_edge

Delete an entity relationship edge by UUID.

clear_graph

Completely clear all data from the knowledge graph (irreversible).