Daniel Chalef

Daniel Chalef is the founder and chief executive of Zep AI and a creator of Graphiti, an open-source framework for building temporal knowledge graphs. His work addresses a central weakness of AI agents: remembering information without understanding whether it is relevant, current or trustworthy.

Originally from South Africa, Chalef founded KnowledgeTree, an enterprise document-management and sales-enablement company, serving as chief executive and chief technology officer. He subsequently led marketing at Domino Data Lab and held senior data-science and corporate-development roles at SparkPost. Those experiences spanned enterprise software, data infrastructure and the practical demands of selling technical products to businesses.

Chalef founded Zep AI in 2023; the company joined Y Combinator’s Winter 2024 batch. His early writing on hybrid search emphasized using metadata and business-specific filters to retrieve useful information from conversations. That approach evolved into Graphiti and a 2025 research paper, coauthored with Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais and Jack Ryan, describing temporal knowledge graphs that combine conversational history with structured business data.

• Business relevance is not semantic similarity. Vector search can mistake a dog named Melody for useful information about music. Chalef’s domain-aware agent memory instead models application-specific entities and relationships: a financial coach can retrieve income, debts and goals using typed schemas, business rules and targeted graph searches.

• Temporal knowledge graphs preserve changing facts. Graphiti tracks when relationships become valid or obsolete, allowing agents to distinguish current circumstances from historical ones. Its retrieval combines graph traversal, full-text search and vector similarity.

• Source provenance makes synthesized facts accountable. A patient’s apparent allergy might originate in a clinical record, laboratory report or unverified intake form. Chalef’s knowledge-graph provenance architecture links derived claims to their source episodes, retains lineage when identities merge and records when newer information invalidates earlier claims.

• Trust and deletion depend on context. An allergy warning might warrant attention when any source supports it, while procedural consent may require independently verified records. Explicit source links also enable selective deletion: a claim survives when other records support it and disappears when its final supporting source is removed. Chalef favors deterministic retrieval and deduplication wherever they can reduce the expense and unpredictability of additional model calls.

His independently maintained mrfparse, a Go parser for healthcare price-transparency files, extends his interest in making complicated real-world information structured and operationally useful.