Rylan Talerico on Zep: A Temporal Knowledge Graph Architecture for Agent Memory [PWL NYC]

We're please to present Rylan Talerico on Zep: A Temporal Knowledge Graph Architecture for Agent Memory - Paper: https://arxiv.org/pdf/2501.13956

----
Papers We Love NYC would like to thank Datadog for making the NY Chapter events possible. Learn more about a career at Datadog: https://careers.datadoghq.com/
---

Today's large language models (LLMs) are stateless: at test time, the scope of their accessible information extends only to their internally encoded knowledge and the tokens in their context window. AI memory has emerged to address this, enabling long-horizon continuity and user personalization in LLM applications by intelligently hydrating the context window before inference. Zep is a low-latency, temporally aware, graph-based AI memory architecture. Zep reports strong performance on the Deep Memory Retrieval (DMR) and LongMemEval (LME) benchmarks, and is among the most well-known architectures in the space today. Zep retrieves and reconstructs relevant information across histories exceeding 115,000 tokens, as demonstrated in LME.

Rylan Talerico is co-founder and CPO of Retriever, where he works on AI memory and personalization. A self-taught engineer who dropped out of high school to self-direct, Rylan founded Crate.fm, a cloud storage and collaboration platform for musicians, before starting Retriever. Outside of work, he loves reading, music, and running.