Changelog

A follow-up question that referred to an earlier answer, such as “when did she move there?”, was searched as written and often returned results that did not match the conversation. The assistant now rewrites a follow-up question into a standalone question before it searches the graph. The panel shows this rewrite as the first step of the turn.

Read the docs

A byom.request_failed webhook event reported the primary provider and model for every failure, also when the fallback provider ran. The event now names the provider and model that produced the failure, and a rejected credential now reports BYOM_CREDENTIALS_INVALID. Ingestion no longer retries a request that a provider rejected because of the credential, so the failure reaches you on the first attempt.

An episode list showed the ingestion time for a message that was added with its own created_at. The list now shows the time you set.

Read the docs

A user summary stayed at its initial text when the user graph changed through fact triples, node adds, or batch ingestion. Zep now rebuilds the summary with the episodes added since the last update.

Read the docs

A thread message ingested through the Batch API did not carry the speaker and the user identity that thread.add_messages adds during extraction, so the same conversation produced a different graph. Both paths now produce the same graph. Reads return the content you sent, and usage is metered on that content.

Read the docs

Enterprise accounts can create pre-registered OAuth clients for Gemini Enterprise from the MCP Server page. The page shows the client secret once and lists and deletes existing clients. Listing and deletion stay available on every plan.

The graph visualizer in the dashboard has an assistant that answers read-only questions about the graph you are viewing. It shows each retrieval step and cites the facts and entities behind the answer. The assistant is in limited rollout, so Ask graph is shown only when it is enabled for your account.

Read the docs

You can now select gemini-3.8-flash for Vertex AI and Gemini in LLM Providers.

Read the docs

Custom edge attributes on a fact triple were cleared on every triple after the first one added to a graph. The attributes are now kept.

Read the docs

Metadata on a thread message now behaves the same as metadata added through graph.add. Zep projects it onto the facts, entities, and summaries derived from the message, graph search filters on it, and source-based access control evaluates it. Message metadata accepts at most 10 keys with scalar or array-of-scalar values. On an update, a key set to null is removed.

Batch ingestion ignored the setting that disables the default ontology, so the same content could receive different entity and edge types when it entered through a batch. Batch ingestion now applies the setting in the same way as graph.add and thread.add_messages.

Read the docs

Extraction dropped a fact whose source and target are the same entity, which lost personal states and routines. Those facts are now kept in the graph.

The debug mode Reasoning tab is now Ingestion Traces. A trace explains why each traced ingestion stage produced its decision. Tracing does not change the stage output or the graph. Zep retains traces for 30 days.

Read the docs

graph.get_episodes_for_document and the thread episode list returned no episodes for documents and threads that had them. Both endpoints now return the episodes, including those ingested before this fix.

Read the docs

An episode that failed during batch ingestion and was retried lost its document_id, so it completed ingestion without document membership and did not appear in the document’s episode list. A retried episode now keeps the document_id and the other ingestion options from the original request.

Read the docs

You can now select gpt-5.6-luna for OpenAI and Azure OpenAI in LLM Providers. Zep sends OpenAI requests for this model with reasoning disabled. An Azure deployment name must match gpt-5.6-luna.

Read the docs

Onboarding offers an MCP server option beside the SDK snippets in the wire-up step. It links the setup guide for each supported host and gives the endpoint for other MCP clients. Enable an MCP connection for the project first.

Read the docs

Opening an account page and then returning to your projects opened the first project in your list, not the project you were working in. The dashboard now remembers the project you viewed last and returns you to it. It falls back to the first project you can access when that project is gone.

A guide on how to keep retrieved context out of privileged model instructions was added to Best Practices.

Read the docs

Guidance on when to give episodes a shared document_id was added to the Documents page.

Read the docs