Preston Rasmussen, Zep AI, Senior Software Engineer

Zep is a novel memory layer for AI agents designed to address the limitations of current RAG frameworks in dynamic, real-world enterprise applications. Unlike traditional systems that rely on static corpora, Zep integrates a temporally-aware knowledge graph engine, Graphiti, to dynamically synthesize unstructured conversational data and structured business data while maintaining historical relationships. Zep’s hierarchical knowledge graph structure mirrors human memory, consisting of episodic, semantic, and community subgraphs that store raw data, extracted entities, and high-level domain summaries, respectively.

Key findings demonstrate Zep's superior performance in memory benchmarks. In the Deep Memory Retrieval (DMR) task, Zep achieved a 94.8% accuracy, outperforming MemGPT (93.4%). On the more demanding LongMemEval benchmark, Zep achieved up to an 18.5% accuracy improvement and reduced latency by 90%, excelling in complex tasks like temporal reasoning and multi-session information synthesis.

Zep’s innovations include a bi-temporal model for accurate historical tracking, efficient retrieval through multi-method search and reranking, and scalable, latency-optimized performance. These advancements make Zep particularly suited for enterprise use cases such as customer support, business intelligence, and decision-making. By enabling real-time integration of evolving data, Zep represents a significant step forward in enhancing LLM-agent memory for real-world applications.