FalkorDB, Cognee and OpenClaw: agent memory tutorial

This webinar explains graph-based retrieval and persistent AI agent memory through FalkorDB's GraphRAG-SDK and a Cognee integration with OpenClaw. The speakers argue that vector search can miss related information when relevance depends on connections between entities rather than similarity to a question.

Gal Shubeli describes an ingestion pipeline that links documents and chunks, extracts entities and relationships, resolves duplicates, and builds indexes. Retrieval combines search with graph traversal. The SDK supports generated or custom ontologies, incremental document updates, and traces that explain retrieval steps. Its database connection can point to a local host or the cloud. A local entity-recognition model handles an initial extraction pass; an LLM then verifies entities and extracts relationships.

Guy Lubovitch presents a prerecorded demo of his Ubuntu laptop setup with OpenClaw, Cognee and FalkorDB, using Telegram to communicate while away. His work and personal agents have separate named memory graphs. He adds information about contacts and companies in Tokyo, then queries the graph browser to inspect their relationships.

The local FalkorDB server and GraphRAG SDK are separate components. Configuring a local database does not by itself keep model-provider requests on your machine. They also state that community detection is absent from the presented GraphRAG solution and that their benchmark tested documentation, not code. The closing discussion introduces code graphs and Graphify's FalkorDB support for a coding assistant.