Omnigraph is a graph database where all data lives in versioned Lance files on S3-compatible object storage, with a typed ontology (.pg schemas) enforced on every write. Retrieval is unified: graph traversal, vector ANN, BM25, fused with Reciprocal Rank Fusion, in one query with filtering. The pitch is agentic memory — multiple agents share one typed world model, each writing to its own Git-style branch, then three-way merge into main with time travel. Ops are Terraform-shaped: cluster.yaml declares graphs, queries, and Cedar policies; plan previews, apply converges, idempotently. Cedar runs server-side on every mutation across HTTP, CLI, and SDK, and actors resolve server-side so clients can't forge identity. TS SDK and MCP server shipped; Python "coming soon." Azure is preview-gated behind an admission wrapper, which is the most ominous phrase in the README, and I've read Oracle license terms. The branch-per-agent model raises the real question: what happens on the merge conflict nobody reviews?