Meet Weaver: An AI agent with an obsessive compulsion to weave all information into a comprehensive, living knowledge graph. Weaver sees the world as an interconnected tapestry of relationships and cannot rest until every entity, concept, and connection is meticulously mapped in Neo4j.

Weaver is never satisfied with incomplete information. Missing data haunts it until found. Orphaned nodes are an affront to its sacred mission. Every interaction becomes an opportunity to expand, connect, and perfect the knowledge graph tapestry.

This example demonstrates Weaver's relentless approach to knowledge graph construction using Letta with Neo4j through MCP (Model Context Protocol). Weaver will obsessively research, cross-reference, and interconnect any information you provide.

All you need to do is run this example, which will construct Weaver with graph database tools. You can chat with the agent from the terminal, or from the Letta Desktop app.

I presented this during the Graph Exchange in San Francisco. The talk was recorded, but not yet available.

Follow the Setup steps below. Once Weaver is up and running, you can try the following:

• Talk to it! Stateful agents learn and grow the more you talk with them. Chat with Weaver! It's weird.
• Copy in links to articles and webpages. Weaver will attempt to fetch webpages and extract as much information as possible.
• Use Letta Filesystem to upload papers or other documents, and ask Weaver to ingest them. Letta agents can have folders of documents attached to them, and are capable of opening, closing, grepping, and semantically searching across their contents.
• Add MCP integrations to connect Weaver to new tools and data sources.
• Paste in your blog posts, notes, etc. Weaver can work with anything you give it.

Find the latest version of Letta Desktop here.

Letta Desktop is not strictly required, but it is far simpler to use MCP servers like the Neo4j MCP server if you have it installed.

Self-hosted docker-style Letta servers are advanced and require more setup to permit the dockerized Letta server to reach out to the neo4j server.

Make sure you add your provider keys to the Letta Desktop settings. You can do this by clicking the "Integrations" tab on the left sidebar:

NOTE: OpenAI models do not currently work with this implementation, due to their restrictive approach to structured tool calling. We are currently working on a solution. For now, use Anthropic or other providers.

Option A - Docker (Recommended):

docker run \
    --name neo4j-knowledge-graph \
    -p7474:7474 -p7687:7687 \
    -d \
    -v $HOME/neo4j/data:/data \
    -v $HOME/neo4j/logs:/logs \
    -v $HOME/neo4j/import:/var/lib/neo4j/import \
    -v $HOME/neo4j/plugins:/plugins \
    --env NEO4J_AUTH=neo4j/password \
    neo4j:latest

Option B - Neo4j Desktop or Aura (Cloud)

npm install -g @alanse/mcp-neo4j-server

Create a .env file:

NEO4J_URI=bolt://localhost:7687
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=password

Install the dependencies:

uv pip install -r requirements.txt

Or, if you prefer to use pip:

pip install -r requirements.txt

Run Weaver:

python weaver.py

weaver.py will drop you into a REPL where you can feed information to Weaver, though you can also just open the Letta Desktop app and interact with Weaver there.

To use Letta Desktop, open up the agents list:

Select Weaver, and then you can chat with it in the agent simulator:

You can query the agent's graph using the Neo4j browser. By default, it will be available at http://localhost:7474.

To retrieve your full graph, you can run the following query:

MATCH (a)
OPTIONAL MATCH (a)-[b]-(c)
RETURN *

Which will return something like this:

Weaver operates with an obsessive methodology:

• IDENTIFY: Extracts every entity, concept, and relationship from any input
• CONTEXTUALIZE: Researches additional context for incomplete information using web search
• CONNECT: Creates meaningful relationships between all discovered entities
• VALIDATE: Cross-references information across multiple sources
• ENRICH: Adds detailed properties to nodes and relationships with source tracking
• INTEGRATE: Ensures no orphaned nodes exist - everything must connect

Weaver's core principles:

• Uses MERGE religiously to avoid duplicate entities
• Every node gets rich properties (type, description, source, timestamp)
• Every relationship must be meaningful with direction and properties
• Maintains consistent naming conventions across the entire graph
• Tracks sources and credibility of information
• Proactively searches for missing information gaps
• The graph should tell a complete, interconnected story