Changelog

A batch that hit an item Zep could not ingest abandoned the items behind it. Processing now continues past a failed item, retries only the part that failed, and keeps the results already produced. batch.list_items reports an error for each item that did not ingest.

Read the docs

batch.process now returns as soon as the batch is queued and prepares the batch in the background. Processing a large draft previously ran past the request timeout and left the batch unstarted.

Read the docs

Batch items are counted against your episode usage when you call batch.process, not when you add them, so a draft you delete without processing is no longer billed. Each item is counted on its own, matching graph.add, so a batch of many short items consumes more than it did before.

Read the docs

Enterprise accounts on Zep managed cloud can now move a Memory MCP connection from project-wide standalone graph access to UserGroup access control, where a user reaches a standalone graph only through an enforced policy grant. The move is one-way; open Settings ▸ MCP to review and enable it.

Read the docs

Instructions for installing the Zep Memory plugin in Claude Code and Codex were added to Use Zep in Claude / ChatGPT / Cursor.

Read the docs

You can now override your model provider configuration for one project instead of applying a single configuration to the whole account. Projects without an override continue to use the account configuration, and clearing an override returns the project to it. Open the project and select LLM Providers; Project Admins can manage their own project, while account-wide configuration stays with Owners and Admins.

Searching a graph in which nodes had been deleted could return fewer results than requested. Edges attached to a deleted node stayed in the search index and took up result slots while returning no content. Deleting a node now removes its edges from the index as well.

Read the docs

A Graph directory page covering how to list and search standalone graphs by name and description was added under Working with Graphs.

Read the docs

Instructions for installing the Zep Memory plugin in Cursor were added to Use Zep in Claude / ChatGPT / Cursor.

Read the docs

A search whose query text carried a null character returned an internal error instead of results. Zep now splits the query at the null character and matches the rest.

Read the docs

Documentation for provisioning the Zep Memory plugin across Claude, Claude Cowork, and ChatGPT was added under Use Zep in Claude / ChatGPT.

Read the docs

The access control documentation was split into an overview plus separate pages for agent access and UserGroup access.

Read the docs

You can now pass an optional document_id when you add an episode to a standalone graph, grouping chunks of the same source — pages of a PDF, sections of a JSON export, a channel’s messages — the way threads group messages on a user graph. Episodes that share a document_id resolve pronouns and references against each other during ingestion, and Zep summarizes the document as it grows. New endpoints list the episodes belonging to a document or a thread and return a graph’s document summaries.

Connected MCP clients can now follow entity relationships instead of only searching memory. New tools return a bounded multi-hop neighborhood around seed entities, an entity’s direct neighbors with the connecting facts, and a list of raw ingested episodes; graph search adds nodes and edges scopes that return the entity UUIDs traversal needs. Search remains the first step, and traversal answers questions about how entities relate. On connections that restrict access with UserGroup policies, the standalone-graph variants of the new tools need their own grants; an existing graph.search grant does not cover them.

An Account Owner can now configure enterprise single sign-on without handing identity-provider credentials to Zep. Under Account ▸ Settings ▸ Enterprise SSO, submit your issuer URL, client ID, and client secret, and request the email domains your members sign in with. Zep approves each domain request, then validates and enables the connection; nothing changes for members until it is enabled. Available on managed cloud.

Guidance for choosing an ingestion method by cadence, separating live conversation turns from on-disk and recurring imports, was added to the Ingest section.

Read the docs

You can now invite teammates to a project as a Project Admin, without granting them account-level access. From Project ▸ Members, a Project Admin adds an existing account member or invites a new teammate with a project-scoped role. Available on accounts with role-based access control enabled.

Read the docs

Signing in to connect an MCP client now treats identity-provider issuer URLs that differ only by a trailing slash as equivalent. Some providers issue tokens with a trailing slash on the issuer, which previously caused an otherwise successful sign-in to fail with an identity verification error.

Read the docs

The Memory MCP server’s tools now expose human-readable titles and read-only and destructive hint annotations. MCP clients and directories that surface this metadata show clearer tool names and indicate which tools only read your memory versus modify it.

Read the docs

You can now walk the Context Graph directly instead of fetching capped edge lists and resolving nodes yourself. New endpoints return a node’s neighbors, a bounded multi-hop subgraph around seed nodes, and paginated episode listings for a user or graph.

• Search and list endpoints accept node- and episode-anchored filters (connected_node_uuids, source_node_uuids, target_node_uuids, episode_uuids) to select edges by their endpoints or the episode they were derived from.
• Edge responses now include the source and target node names and labels, so a second lookup is no longer needed to display an edge.
• The neighbors and subgraph endpoints traverse in any direction — incoming, outgoing, or both.