Open-source AI avatar adds knowledge-graph memory

What happened

A developer published an open-source AI avatar that uses a knowledge graph as its memory system, allowing real-time conversation. The system combines vector search speed with knowledge-graph context, powered by Zep's open-source knowledge retrieval framework Graphiti (available under Apache 2.0). The setup uses fine-tuned Qwen3 and Gemma models with embedding and reranking times under 10ms and 50ms respectively, plus S3 with hot caching for dense vector and BM25 search.

Why it matters

Traditional RAG (retrieval-augmented generation) systems treat documents as isolated chunks without context, while knowledge graphs are smart but slow in real-time use. This project demonstrates a working middle ground—combining the speed of vector search with the contextual understanding of knowledge graphs—showing a practical way to give AI agents richer memory without sacrificing response time. The code and framework are fully open-sourced, so anyone can self-host it.

What to watch

Zep's Graphiti framework is available under Apache 2.0 and the full code is open-sourced on GitHub, making this immediately accessible to developers who want to build similar memory-aware AI systems locally without reliance on proprietary databases.