Graph RAG

Graph RAG in Meko uses Apache AGE (A Graph Extension) to store and query entity-relationship data extracted from agent conversations. This complements the vector-based memory by providing structured relationship traversal.

How it works

When memories are added, Meko's mem0 integration extracts entities and their relationships:

1. An LLM analyzes the text to identify entities (people, places, concepts, preferences) and relationships between them.
2. Entities become nodes in the graph, stored in Apache AGE.
3. Relationships become edges connecting those nodes.
4. A similarity threshold (default 0.7) prevents duplicate nodes for semantically equivalent entities.

For example, the memory "I like tropical beach vacations on a budget" might create:

• Nodes: User, Tropical Beaches, Budget Travel
• Edges: User --prefers--> Tropical Beaches, User --prefers--> Budget Travel

Why graph + vector

Vector search alone finds memories that are semantically similar to a query. Graph traversal adds the ability to:

• Follow relationships: "What else does this user prefer?" can traverse edges from the user node.
• Discover indirect connections: Two memories might not be similar by embedding, but connected through shared entities.
• Provide structured context: Graph edges carry relationship types, giving the LLM more structured context than raw text similarity.

Meko's memory search combines results from both pgvector (semantic) and Apache AGE (graph) to provide comprehensive recall.

Configuration

The graph store is configured with connection details to the Apache AGE instance and a similarity threshold for node matching: