When an agent answers from a context graph, it's hard to see which facts it used. Our graph assistant in the Zep dashboard shows each search it runs and the evidence behind the answer. We launched Zep v4 this week with a lot of new features. See them all below.

We're excited to announce the launch of the Zep v4 API! Agents can now navigate a graph by connection. They see how each episode was built using Ingestion Traces and use hyperedges to retrieve facts that connect more than two entities. Launch post below.

Your agents don't share what they know. Claude and ChatGPT can't reach company context, and the agents you build never see what those assistants learn. Zep's Memory MCP Server ships today: one agent memory across all of them. 🧵

Agent memory is synthesized: an LLM reads chat history, documents, and business data, and derives the facts your agent retrieves. The output appears in none of the sources verbatim. By retrieval time, the source behind a fact is gone. Here's how Zep keeps it. 🧵

Your team's agents each keep their own agent memory: Claude Cowork, Codex, Code, your custom agents. They all model the same problems and users separately. Zep's Memory MCP Server puts them on one governed user graph. 🧵

Why it matters: ⚡ Faster than ingesting items one at a time 📦 Up to 50,000 items per batch 🔌 Runs separately from real-time ingestion 👀 Track status, progress, and per-item errors in a new dashboard

Agents need more than one kind of context to actually help users. Zep produces 5 distinct types in every Context Block: facts, entities, episodes, thread summaries, and the user summary. Each captures something the others can't. Here's what each is for 🧵

When your agent acts on bad context, the first question is: where did that come from? Most context systems don't track this. Zep does, enabling: → Audit trails from response to source → Inspectable debugging paths → Agents scoping their own context by source Here's how 🧵