[Hands-on] Build Your Own AI Avatar With Human-like Memory

A hands-on guide to building a real-time AI avatar with human-like memory using a knowledge graph backend. The solution combines vector search speed with knowledge graph intelligence via Zep's open-source Graphiti framework (Apache 2.0), using fine-tuned Qwen3 and Gemma models for fast embeddings and reranking, plus S3 hot caching instead of traditional vector databases. The post also recaps 9 MCP (Model Context Protocol) projects covered previously, including local MCP clients, agentic RAG, voice agents, financial analysts, and shared memory layers for Claude Desktop and Cursor.