Summary
Daniel Chalef, co-founder of Zep AI, presents at AI Engineer World’s Fair on engineering provenance and lineage into LLM-built knowledge graphs. The core problem: when LLMs synthesize facts from multiple sources, the paper trail is destroyed — output artifacts don’t appear verbatim in any single source, and tracking “where did this fact come from?” breaks down with standard source-ID tagging or append-only logging when entities merge, facts mutate, and the graph keeps changing underneath any pointer.
Chalef’s team built Graphiti, an open-source temporal graph framework that underlies Zep’s enterprise agent memory infrastructure. Rather than logging provenance after construction, Graphiti bakes lineage into the graph structure itself: source episodes link directly to extracted entities and edges, and those links survive mutations. The talk demonstrates metadata projection — classification tags applied at ingestion propagate through all derived facts — enabling veracity filtering critical in healthcare scenarios where an allergy flagged only in a patient’s self-reported intake form should be treated differently than one from a verified EHR record.
The framework also handles GDPR-style data deletion: a fact survives only if at least one supporting source episode remains, making compliance-driven deletion tractable even when a fact has multiple parents. Chalef closes by noting that graph construction is expensive, and Zep has invested heavily in reducing cost and latency. The open-source Graphiti repository on GitHub contains most of what was covered.
📺 Source: AI Engineer · Published July 23, 2026
🏷️ Format: Deep Dive
