Changelog

When an entity name contains a different named entity, Zep now also extracts the named entity as a separate entity. For example, the text “Acme Cloud billing outage” now gives the entity “Acme Cloud” and the entity “Acme Cloud billing outage”. A fact can then connect the two entities. Zep does not split a proper name that identifies one entity, such as “Bank of America”.

The agent.skill.search method now accepts a query that has more than 64 terms. Search uses the first 64 terms of the query and ignores the other terms. Before, a query with more than 64 terms returned an HTTP 400 error. The maximum query length of 512 bytes does not change.

You can now group custom edge types into types and subtypes with the optional edge_type_hierarchy field on project.set_ontology and graph.set_ontology. Zep classifies each extracted edge into the deepest subtype that applies. A search filter on an edge type also matches its subtypes. This feature is in closed beta and is available only for select accounts.

The Graph Assistant is now off for a HIPAA account until the account opts in. To opt in, an account owner or admin turns on Graph Assistant on the Account Settings page of the dashboard. Before the opt-in, the assistant sidebar shows a message that tells you how to opt in.

Zep now continues to update the observations of a graph when the model returns an observation with an empty name. Before, this answer stopped all observation updates for that graph.

When a user has no user ID, the user node in the user graph now uses the user UUID as its identifier. Before, the user node summary showed an empty identifier.

You can no longer select GPT-5.4 Mini or GPT-5.4 Nano as the model of a BYOM configuration. The change applies to the OpenAI provider and to the Azure OpenAI provider. Select GPT-6 Luna or GPT-5.6 Luna in place of these models.

Guidance to use the Batch API for test data and backfills was added to the Adding Context section.

The v4 graph.node.add and graph.edge.add methods now validate each custom attribute against your ontology. A node with attributes must have a label that matches an entity type. An edge with attributes must have a fact_name that matches an edge type. An attribute that the matching type does not define, or a value of the wrong type, returns an HTTP 400 error. Before, Zep accepted these attributes without validation.

Hyperedges and entity type hierarchies are now on by default on BYOC. Each new project on a BYOC deployment uses hyperedges. To use hyperedges in an existing project, set hyperedges_enabled to true on the project.

Documentation of the batch.process task status for each batch status was added to the Batch API guide.

Documentation of how a graph ontology uses the project entity types or edge types was added to the Customizing Graph Structure guide.

You can now filter the list of tasks by status and by task type in the v4 SDKs for Python, TypeScript, and Go. The filters are available in the v4 SDKs at version 4.0.1.

You can now select Claude Haiku 5.5 as the model of a Bring Your Own Model (BYOM) configuration. The model is available for the Anthropic provider and for the Amazon Bedrock Anthropic provider.

A BYOM configuration now stays active when you change the provider type in the dashboard. Before, Zep could use its own models for some requests during the change. A change of the provider type without credentials now shows an error and does not change the configuration.

Content policies now try to repair an artifact that matches a rule. Zep removes the prohibited meaning from the artifact and evaluates the repaired text again. Zep publishes the repaired artifact when no rule matches it. Before, Zep dropped each artifact that matched a rule.

Sign-in to the Context MCP Server with Custom OIDC now works when your identity provider application requires PKCE. An example is an Okta application with “Require PKCE as additional verification” set.

Scope instructions documentation was added to the Custom instructions page.

Documentation for the zepctl retention target commands was added to the Context retention page.

The graph visualizer assistant is now available to all accounts. Before, the assistant was in limited rollout, and the visualizer showed Ask graph only for some accounts.