Zep bets AI-agent memory is temporal, open-sources Graphiti; 30.6k stars

Zep is an AI-agent infrastructure company whose bet is that agent memory is a temporal problem. Traditional RAG flattens documents into chunks and summaries, but what an agent needs to remember changes continuously: Kendra loved Adidas shoes in March 2026 and may not in June. Graphiti, the framework Zep open-sourced, treats every fact as having a validity window — when it became true and when it was superseded — keeps full provenance back to the raw episodes that produced it, and updates the graph incrementally as new data arrives.

Graphiti launched on Hacker News as a Show HN on 2024-09-04 (142 points, 21 comments). The pitch was deliberately anti-RAG: hybrid retrieval combining semantic embeddings, keyword search and graph traversal, no full-graph recomputation on ingest, prescribed or learned ontology, and sub-second historical queries. Zep positioned the framework as the open core of its own commercial platform, which manages context graphs at scale on a proprietary Context Graph Engine with sub-200ms retrieval, governance, dashboards and SDKs.

Adoption compounded through 2026: the GitHub repo reached 30.6k stars, 3.1k forks and 965 commits by the 2026-09-04 crawl, and the README's Zep-vs-Graphiti comparison shows the commercial mechanism — self-host the framework if you want flexibility, buy Zep if you want it operated. The January 2025 architecture paper formalized the design, and the repo's MCP server and FastAPI service kept the framework easy for agent developers to adopt.