Zep and Graphiti: agent memory architecture vs RAG

Preston Rasmussen, a software engineer at Zep AI, explains how Zep approaches AI agent memory with Graphiti, its open source graph framework. He contrasts continuously updated, user-specific memory with RAG systems built around largely static reference documents. The talk covers architecture and retrieval rather than installation.

Graphiti stores raw messages and structured data as episode nodes, then uses LLM calls to extract entities and relationships. Community summaries provide broader context. Its temporal model tracks both when events happened and when the system learned about them. Changed facts can become invalid without disappearing, so retrieval can distinguish current preferences from past ones.

Search combines vector similarity, full text matching and breadth-first graph traversal before reranking results. Rasmussen describes a Neo4j backend and says the team self-hosts its indexes. This does not establish an offline deployment or local model configuration.

Rasmussen reports better LongMemEval results than passing the full conversation context to an LLM. These are the team's reported results. Practical limits also receive attention: callers handle document chunking, community assignments can drift, and he recommends retaining three to five recent conversation turns for short-term context.