Changelog

Debug mode documentation — capturing per-episode workflow logs and Enterprise-only reflection traces from the dashboard — was added to the Key Concepts section.

Read the docs

You can now create up to 100 entity nodes in a single call with graph.add_nodes. Unlike ingestion, you supply the node contents directly and Zep skips extraction. Client-supplied UUIDs are supported for upserts and safe retries.

Read the docs

MCP connections now default to Zep-operated social login — users sign in with Google, Google Workspace, GitHub, or Microsoft. Enterprise plans can still connect a customer-owned IdP through the brokered OIDC flow.

Read the docs

Calling batch.process on a batch that already has a run now returns a conflict instead of silently starting a second run; concurrent calls collapse to a single run.

Read the docs

Ontology validation errors now name the offending category and its actual count — for example, entity_types cannot contain more than 10 items (got 11) — instead of a single combined message.

Read the docs

Attribute-Based Access Control (ABAC) — scoping an API key to a subset of a project’s actions and data — is now fully documented.

Read the docs

A new page explains how an episode’s metadata projects onto the facts, entities, and summaries derived from it — the basis for metadata filtering in search and source-based ABAC policies.

Read the docs

Guidance on avoiding overlapping custom entity and edge type definitions — types are matched on their descriptions, not their names — was added to the Customizing graph structure page.

Read the docs

The Memory MCP Server now exposes three tools — search_graph, get_user_summary, and add_memory — replacing the previous eight. search_graph returns Zep’s assembled, reranked context block by default. This is a breaking change for MCP clients that used the removed tools; the REST API and SDKs are unaffected.

Read the docs

The Memory MCP Server (Enterprise, enabled per account) is now fully documented — setup, authentication, and connecting a client.

Read the docs

New guides cover AG2, Mastra, Microsoft Agent Framework, Pydantic AI, and the Vercel AI SDK, and the existing framework guides have been refreshed.

Read the docs

The ontology view now shows the source and target restrictions (Valid Connections) configured on each edge type.

Read the docs

Updating a graph edge with graph.edge.update now clears the valid_at and invalid_at timestamps when you set them to null, and attribute updates replace rather than merge — so an attribute set to null is removed. Cleared values are reflected in later searches.

Read the docs

Thread objects returned by thread.create, user.get_threads, and thread.list_all now carry the thread’s actual project_uuid instead of an all-zeros UUID.

Read the docs

The analytics dashboard now accepts an arbitrary start and end date in addition to the 24h, 7d, 30d, and 90d presets. The available lookback follows your plan — up to 7 days on non-Enterprise plans and 90 days on Enterprise. In the dashboard, open Analytics and pick a custom range.

The dashboard’s “View Ontology” dialog is now a “View Customization” dialog with tabs for custom ontology, custom instructions, and user summary instructions, available on the graph, user, and project settings pages. Each tab states whether the shown configuration is project-wide or specific to that graph or user. Custom instructions remain limited to Enterprise and Flex Plus.

The graph visualization now uses a size-scaled layout that spreads dense clusters apart, so nodes and their connections are easier to read.

The metadata you attach to an episode now resolves onto the facts, entities, and summaries derived from it, and is returned when you retrieve those objects — previously this metadata could be filtered on but not read back. The metadata of every contributing episode is included, and numeric metadata values now match filters consistently.

Read the docs

You can now use the Batch API on any plan, not only Enterprise. Authenticated accounts with the appropriate API key permissions can create batches, add items, and process them through the /batches endpoints. Existing entitlement and rate-limit checks still apply.

Read the docs