Steer observation generation (experimental)
You can now guide how Zep generates observations by setting an instruction and up to ten custom observation types, scoped to your project or to a specific user or graph. Steering applies to observations generated after you configure it, and the observation type Zep selects is searchable on the graph.
Graph search now respects property filters
Graph search now correctly applies the node and edge property filters you pass, so
results are limited to entities and relationships matching every requested filter.
Previously these filters could be dropped, returning results that did not match
conditions such as status = done.
Reasoning traces in debug mode for Flex and Flex Plus
Debug mode now includes reasoning traces on the Flex and Flex Plus plans, not just Enterprise, so you can review the model’s reasoning when inspecting ingestion output. The debug control is disabled on ineligible plans, and Enterprise keeps its longer debug window. The output previously labeled “reflection” is now called “reasoning”.
Manual graph updates documentation
Documentation for manually updating the graph — creating and updating nodes and edges, alongside fact triplets — was expanded under Working with graphs, with Python, TypeScript, and Go examples.