
Research from the Zep team
The Zep team studies how agents keep up with changing business facts and learn from their own runs. Both papers below ship in Zep today.
- November 2026 · arXiv 2611.01386Ships as Zep Agent Skills
Learning by Doing: Autonomously Compiling and Curating Agent Skills from Verified Traces
Daniel Chalef, Peter Evans, Pavlo Paliychuk, Mehul Patel, Preston Rasmussen, Jack Ryan
Most agent memory covers facts and past conversations, and research systems that learn procedures from agent runs have mostly stayed on benchmarks. Zep Agent Skills brings procedural memory to production agents by compiling their verified runs into versioned skills that cite the events they came from. Zep checks every version before agents can load it, and the skills use the open Agent Skills format that any compatible agent can read. In the paper’s tests, smaller models that loaded skills built from a stronger model’s runs completed more tasks at every reasoning effort, with no retraining.
- January 2025 · arXiv 2501.13956Runs in Graphiti and Zep
Zep: A Temporal Knowledge Graph Architecture for Agent Memory
Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais, Jack Ryan, Daniel Chalef
Business facts change, and retrieval built for static documents loses track of what is true now and what was true last quarter. This paper introduced the temporal context graph architecture behind Graphiti and Zep, in which every fact keeps two timelines, when it was true and when Zep learned it. When new information contradicts a fact, Zep closes the old fact and keeps it on record, so agents can answer for the present or for any date in the past. One graph holds both conversations and business data, and Zep searches it by meaning, keywords and graph distance before assembling the agent’s context.


