Zep - A Temporal Knowledge Graph Architecture for Agent Memory

Explore a temporal knowledge graph architecture designed to address the stateless nature of large language models through this conference talk from Papers We Love NYC. Learn how Zep enables long-horizon continuity and user personalization in LLM applications by intelligently managing context windows before inference. Discover the low-latency, temporally aware, graph-based memory system that demonstrates strong performance on Deep Memory Retrieval and LongMemEval benchmarks, capable of retrieving and reconstructing relevant information across histories exceeding 115,000 tokens. Gain insights into how this architecture tackles the fundamental limitation of LLMs having access only to their internally encoded knowledge and current context window tokens, and understand the practical applications of AI memory systems in modern language model deployments.