Changelog

A string identity value now matches an existing entity when the two values differ only in case or in whitespace. For example, CITIZENS FINANCIAL GROUP, INC. and Citizens Financial Group, Inc. match. The matched entity keeps its stored value. Before, such values did not match, and the graph could get duplicate entities. The Prepare data for ingestion page now describes this comparison.

Zep now merges two entities with the same or a similar name without a model check only when the entities have a shared neighbor. Other name matches go to the model-based resolution step. Before, a generic name, such as “auto loan”, could merge the entities of two different owners into one node.

Graph extraction now uses custom instructions only for Flex Plus and Enterprise accounts. After a downgrade to a different plan, Zep keeps the stored instructions and uses them again after an upgrade. You can list and delete custom instructions on every plan.

Zep now processes the messages of a thread when a different graph in the project cannot be read. Before, in rare cases, the processing of the messages of a new thread failed for this reason.

The list_episodes, list_episodes_in, get_subgraph, and get_node_neighbors tools of the Context MCP server now return results that MCP clients accept. Before, the output schemas of these tools declared each UUID as an array of bytes, and MCP clients rejected the valid results.

A later episode no longer changes a non-empty identity attribute value on an existing node. Identity attributes are the attributes that an entity type lists in identity_properties. Extraction fills an identity attribute only when the stored value is empty. Before, a later episode could replace a seeded identity value. A note about this behavior was added to the Prepare data for ingestion page.

Batches for the same graph now start in the order of their batch.process calls, also after a retry. Before, a later batch could start before an earlier batch on the same graph. In rare cases, a later batch can still start first. Batches for different graphs have no order.

Zep now uses the identity properties of an entity type to find an existing entity in all graphs. Before, in some graphs, an identity property match occurred only when a name search or an embedding search also found the existing entity. Such graphs could get duplicate entities.

Zep now processes the episodes of a user whose first name or last name contains only spaces. Before, the processing of each episode of such a user failed. If the first name and the last name are both blank, Zep uses the user ID as the name of the user node.

You can now send a node summary of up to 5,000 characters to graph.node.update. Before, the limit was 500 characters, and you could not write back a summary that Zep generated.

An account can now have more than one Account Owner. When an Account Owner gives the Account Owner role to a member, the current Account Owners keep their role. You cannot remove the last Account Owner of an account.

The dashboard now shows credit usage to all Zep Cloud accounts. The Billing page in the EU region also shows credit usage.

During an Enterprise trial, you can now select a self-serve plan on the Billing page of the dashboard.

Auto search now uses the reranker when a search has more than 128 candidates. The reranker scores up to 200 candidates. Before, such a search used reciprocal rank fusion (RRF) order and did not use the reranker.

A guide to build an agent that uses Zep through tools was added in the Build an Agent with Zep section.

Each call to the deprecated graph.add_batch and thread.add_messages_batch methods now creates a Batch API batch. The batch shows in batch.list and in the dashboard. Zep sends the ingest.batch.completed webhook for these batches on the same batch outcomes as other Batch API batches. The task status that the call returns follows the batch status.

The minimum interval between two summary updates for one entity is now 15 minutes. Before, the interval was 30 minutes.

In the dashboard graph visualization, each edge label now has an opaque background with a border in the edge color. The labels are clear in dark mode and in light mode. A label longer than 32 characters is shortened.

This change applies to a graph with more than 1,000 episodes. When Zep extracts facts from a new episode, Zep now uses the most recent episodes as context. Before, Zep could use older episodes as context.

Each custom edge type can now have up to 100 source-target pairs. The previous limit was 10.