← back
Stop Using RAG as Memory — Daniel Chalef, Zep
Takeaway
Model agent memory after your business domain using a typed knowledge graph instead of dumping facts into a vector store.
Summary
- Zep CEO argues semantic similarity is not business relevance — dumping arbitrary facts into a vector store produces irrelevant recalls (e.g., 'my dog Melody' surfacing for music queries).
- Solution: domain-aware memory via Zep's Graphiti temporal knowledge graph, with developer-defined entity/edge schemas in Pydantic, Zod, or Go structs.
- Demo: a finance-coach app captures explicit business objects (FinancialGoal, DebtAccount, IncomeSource) with field-level validation rules and concurrent multi-search filtered by node type.
- Memory entities, edges, and ontology are registered at app start so Zep automatically slots conversation facts into the right business schema.
- Frees developers from writing custom fact extractors while keeping retrieval narrowly scoped to business-relevant slots.
memoryknowledge-graphzep
Original description
RAG is great for static knowledge retrieval—but terrible at memory. Vectorstore-based systems sold as memory lack relational and temporal awareness, leading agents astray with outdated or ambiguous information. Discover how temporally-aware knowledge graphs—built by the open-source Graphiti framework—solve these limitations. You’ll learn practical strategies to maintain precise, context-rich memory, enabling agents to reason accurately about historical context and knowledge provenance. About Daniel Chalef I’m Daniel Chalef, an engineer turned startup founder currently building Zep, where we're creating AI's foundational memory layer powered by knowledge graphs. Our vision is a world where AI agents reliably handle personalized tasks, from the mundane to the monumental, always prioritizing privacy and compliance. Previously, I've led ML and data science teams, marketing, and corporate development at both early-stage startups and late-stage companies, building data-driven products at scale. My first startup was an open source document management application, KnowledgeTree. When I’m not building Zep (which is seldom 🙂), you’ll likely find me cycling or hiking around the Bay Area with my dog. Recorded at the AI Engineer World's Fair in San Francisco. Stay up to date on our upcoming events and content by joining our newsletter here: https://www.ai.engineer/newsletter