Genspark launches sb-git with agent-native Git, diff, and 1 GB free storage

• Genspark shipped genspark_ai's launch post for sb-git, a hosted Git server aimed at agent memory, with clone, push, branch, diff, blame, rollback, and audit semantics instead of chat-log persistence.
• The official sb-git intro page frames the product as a real Git repo on Genspark, with one token shared across Git, REST, and the gsk CLI, which matches genspark_ai's thread about standard Git access from terminals, Claude Code, and OpenClaw.
• Storage starts at 1 GB for free users, while genspark_ai's launch post says Plus and Pro accounts get 10 GB.
• Genspark's own comparison chart in the attached product graphic positions sb-git as a fix for six common agent-state problems: lost memory, buried history, silent overwrites, siloed sharing, hard-to-undo mistakes, and weak auditability.

You can browse the product intro, grab the CLI package on npm, and the official walkthrough gets unusually concrete about how this works: agents can either use normal git clone and git push flows or skip the working tree entirely with one-shot gsk sb-git cat and gsk sb-git commit calls. genspark_ai's follow-up post also makes the positioning blunt: GitHub is for human collaboration, sb-git is where Genspark thinks agents should live.

The interesting part is not that Genspark added another memory store. It chose Git as the storage primitive.

According to the official intro, sb-git exposes smart HTTP, basic auth, and standard refs, so the same repo can be cloned locally, inspected in a browser timeline, or written to by an agent tool call. The page's demo flow uses gsk sb-git init, then a normal git clone, git commit, and git push sequence.

The built-in command set is small and specific:

• init
• ls
• cat
• commit
• log
• diff
• clone-url
• list-repos

That split matters because Genspark is covering both modes. Full working-tree Git for longer runs, pointwise HTTP-backed CLI calls for agents that only need to read or write a file.

Genspark's launch graphic reduces the pitch to six before-and-after claims:

1. Memory: lost between sessions → persistent and versioned
2. History: buried in chat logs → diffable timeline
3. Collaboration: silent overwrites → safe parallel agents
4. Sharing: siloed per agent → shared across agents
5. Mistakes: hard to undo → one-step rollback
6. Audit: opaque → fully traceable

That is a more opinionated pitch than generic "agent memory." Genspark is treating state as a shared repo with version control, not as per-agent scratch space.

The launch is live now, but the official page is still labeled early access.

Install starts through @genspark/cli on npm, which requires Node.js 18+, then gsk login writes credentials for later Git and CLI use. Genspark also said in genspark_ai's follow-up post that sb-git is accessible through Genspark CLI, and its show teaser says the company plans to demo recent launches, including a new model selector and Microsoft Office integration, on its May 5 Genspark Shipped stream.