Git semantics are landing in databases. Not as a metaphor — as actual branch, commit, merge, and push operations on your data. Five things shipped in the last six weeks that show this is becoming a real category.
DoltLite: SQLite with git operations built in
DoltLite is a drop-in SQLite replacement that exposes version control as SQL functions: dolt_commit, dolt_branch, dolt_merge. The C API is identical to SQLite — swap libdoltlite for libsqlite3 and nothing else changes. If you are embedding a database into an agent or local-first app, you now get commit history and branch isolation without a server. Most sysbench metrics land within 2x of SQLite; worst case is 4x.
DoltLite adds remote sync
April 9 brought clone, push, pull, and fetch against file and HTTP remotes, plus the ability to attach existing SQLite databases. The pattern this completes: an agent writes to a local DoltLite database, commits after each task, and pushes to a remote when done. You get a full audit trail of every state the agent touched, with branching for parallel runs.
Dolt at 160 concurrent agents
Gas Town, an agent orchestrator built by Steve Yegge, uses Dolt for shared task state. Concurrent agent writes corrupt naive shared state; Dolt resolves those conflicts the same way Git resolves merge conflicts. Without Dolt, Gas Town fell over past four concurrent agents. With it, they run 160 on a single host and 600 in Kubernetes.
Cloudflare Artifacts: git storage at machine scale
Cloudflare Artifacts entered private beta on April 16, public beta expected in May. It is Git-compatible versioned storage designed for millions of repos per account — one per agent session if you want. An agent gets an authenticated HTTPS remote and pushes and pulls like any Git client. Pricing is per-operation at $0.15 per 1,000, so throwaway repos are cheap. ArtifactFS is open-sourced.
LanceDB: git branching for vector tables
LanceDB shipped physical branch isolation for vector tables in February 2026, with shallow clones and native Hugging Face Hub integration alongside it. A branch gets its own directory — experimental writes cannot corrupt the production table. Shallow clones reference source data without copying it. The Hub integration lets you publish embeddings and indexes as a single queryable artifact. If you are versioning embeddings across model iterations, this is the tooling that was missing.
The pattern across all five: agents need the same guarantees developers have had for decades — rollback, isolation, auditability, and collaboration. Git gave that to code. These tools are giving it to data. If you are building AI features and wondering how to make your data layer production-safe, talk to us.