SB-Git — Memory your agents can keep

What changes when your agents have memory they can trust.

Six dimensions. The same agent, with and without a versioned knowledge layer underneath.

Dimension

Agents alone

Agents + SB-Git

Memory

✕Lost between sessions

✓Persistent & versioned

History

✕Buried in chat logs

✓Diffable timeline

Collaboration

✕Silent overwrites

✓Safe parallel agents

Sharing

✕Siloed per agent

✓Shared across agents

Mistakes

✕Hard to undo

✓One-step rollback

Audit

✕Opaque

✓Fully traceable

From "save it to SB-Git" to a clone on your laptop.

Four moments. No orchestrator code. The agent uses the same Git you do — plus a small CLI for tool calls that don't want a working tree.

ACT 01

The agent saves its work to SB-Git

You name the tool. The agent reaches for it directly.

You

Save this to a new SB-Git repo acme-research — tomorrow I'll resume from another machine.

Agent

bash · agent

# create the repo, then ask gsk for a ready-to-paste clone snippet
gsk sb-git init -r acme-research
eval "$(gsk sb-git clone-url -r acme-research | jq -r '.commands | join(\"\\n\")')"
cd acme-research

# write notes as files, commit, push — standard git
cat > hypotheses.md <<'EOF'
1. ACME ships in Q3 (per 10-K p.42)
2. Pricing tier B at $49/mo (unconfirmed)
EOF
git add . && git commit -m "day 1: hypotheses" && git push

Agent

For one-off reads or writes inside a tool call I can skip the clone — gsk sb-git cat and gsk sb-git commit hit the same repo with one HTTP call. Same token, same refs.

ACT 02

A second agent reads one file — no clone

An hour later, a different agent on a different VM needs just one file. It skips the working tree entirely and goes pointwise via the CLI.

You

Look at hypotheses.md in acme-research and tell me which ones are still unconfirmed.

Agent B

tool call · gsk

# list the repo, then read the one file I need — zero clones
gsk sb-git ls -r acme-research
gsk sb-git cat -r acme-research -p hypotheses.md

# found tier-B pricing was the open one — record the answer
gsk sb-git commit -r acme-research -m "answer: tier B unconfirmed" \
  --files '{"answers/q1.md":"Tier B at $49/mo is still unconfirmed."}'

Agent B

Three HTTP calls, no working tree, no temp files. The same $GSK_API_KEY authenticates everything; the same refs you'd see from git log.

ACT 03

You pick up on a laptop

Same repo. No agent runtime. Just plain Git over HTTPS.

You

Next morning, train ride. Laptop only. I want to read what the agent wrote, scratch a few notes, and push back.

You

bash · laptop

$ gsk sb-git clone-url -r acme-research
# prints the GSK_HOST line + git clone command, ready to paste

$ git log --oneline
a3f92c1 day 1: hypotheses        — agent · claw-vm
e0c10a2 init

You

Add a line to todo.md, git push. Tomorrow's agent run picks up where you left off.

ACT 04

You inspect in the browser

Open genspark.ai/sb-git. See the timeline of what your agent has been thinking.

https://genspark.ai/sb-git/me/acme-research

me / acme-research · main

2f81a4cadd comp pricing notes — you · laptopjust now

b71d04erevise H2 after reading 10-K — agent · claw-vm2h ago

a3f92c1day 1: hypotheses — agent · claw-vmyesterday

e0c10a2init — agent · claw-vmyesterday

You

Click any commit to see the diff. Click hypotheses.md to read what the agent wrote. Roll back if it went sideways.

Three things make the story above just work.

Real Git underneath

Smart-HTTP, basic auth, standard refs. Clone with git, inspect with libgit2, run any community git skill. No proprietary SDK to learn.

A small CLI for tool calls

gsk sb-git covers the pointwise actions agents do most — init, ls, cat, commit, log, diff, clone-url, list-repos. One HTTP call each. No working tree required.

One token, three surfaces

Your gsk- API key authenticates git, REST, and the gsk CLI. Same identity, same repo, no extra setup.

Four commands to a real, versioned memory store.

Run these on a Claw VM, your laptop, or anywhere git is installed.

STEP 01

Get your API token

Install the CLI and sign in once. Your token is reused by every surface — git, REST, and the CLI.

bash

$ npm i -g @genspark/cli
$ gsk login
# browser opens, you authorize, token is written to ~/.gsk/credentials

STEP 02

Create your first repo

One CLI call. Empty repo, ready to clone.

bash

$ gsk sb-git init -r my-notes

STEP 03

Clone, edit, push — the normal way

Ask the CLI for a ready-to-paste clone snippet, then it's plain Git from there. This is how you and your agent will spend most of your time.

bash

$ gsk sb-git clone-url -r my-notes
# prints { url, commands: [...] } — paste the commands, you're cloned

$ cd my-notes && echo "## Q3 goals" >> goals.md
$ git add . && git commit -m "Q3 goals" && git push

STEP 04

Or: bulk-import an existing folder

Got a directory of notes, transcripts, or session history? Push it whole. Any community git skill that speaks smart-HTTP works the same way.

bash

$ gsk sb-git init -r meetings
$ cd ~/notes/meetings   # 200 files, no problem
$ git init && git add . && git commit -m "import"
$ eval "$(gsk sb-git clone-url -r meetings | jq -r '.commands[0]')"
$ git remote add origin https://x-access-token:$GSK_API_KEY@$GSK_HOST/sb-git/me/meetings.git
$ git push -u origin main