Zep Q1 Product Round-up

In Q1, we shipped updates to Zep focused on improving developer control over knowledge graphs, a significant improvement to fact and graph quality, and overall platform usability. The highlights include a Graph Explorer for visualizing user and group knowledge graphs directly in the web app, a redesigned Playground emphasizing our graph-based memory architecture, and a practical Cookbook featuring concrete examples for customizing Zep's functionality.

Below, we'll break down these updates and new features in more depth.

Zep is the State of the Art in Agent Memory

In research we published in January, Zep outperformed MemGPT on the Deep Memory Retrieval (DMR) benchmark, achieving a 94.8% accuracy compared to MemGPT’s 93.4%; however, the study noted DMR’s limitations in assessing practical performance due to advancements in LLMs.

More importantly, Zep demonstrated stronger results on the LongMemEval benchmark, which aligns more closely with real-world enterprise scenarios, delivering aggregate accuracy improvements of up to 18.5%, with some individual cases surpassing 100%, and simultaneously reducing response latency by 90%.

Graph Explorer

You can now visualize and explore every knowledge graph you create, by visiting the page of the relevant user or group in the web app.

New Playground

We released a new version of the Playground that focuses on our new graph-based memory design. The new playground allows you to:

• Explore the knowledge graphs from preloaded datasets
• Add arbitrary message, JSON, or text data to these data sets and see how the graph changes
• See how the graph, the memory context string, and an LLM's answer to a question change in response to new data

Cookbooks

In our new Cookbook, you can find guided examples of techniques you can use to fully utilize Zep, such as customizing your memory context string or sharing memory across users using group graphs.

Other Updates

Other additions include:

• Enhanced data ingestion: Upgraded models provide higher-quality data ingestion into Zep knowledge graphs.
• ignore_role option: Exclude assistant (or other role) messages from graph ingestion for cleaner, more focused graphs.
• Delete episode method: Easily remove specific Episodes from your graph.
• Add fact triple method: Directly insert structured data into graphs using node-edge-node relationships.
• Get user node method: A convenience method for easily returning the User node in a User graph
• Returning the created Episode UUID: The graph.add method now returns the newly created episode UUID for immediate use.