The Future of Git.
Version-control your AI assistant's context. Diff sessions. Roll back mistakes. Sync to your team.
Works with the AI tools you already use
The Problem
Your AI starts from zero. Every. Single. Session.
You teach it everything
Your AI learns your architecture, your team's conventions, the bugs you've already solved. It gets smart. The session is productive.
Session ends.
Context gone. The AI assistant's memory evaporates. Everything it learned about your codebase — gone.
You explain it again.
Re-architect in chat. Re-establish your conventions. Redo the same debugging explanations. Every. Single. Session.
“It's 2005 — the year before git — for AI context management. Nobody has built the version control layer yet.”
How it works
Every session. Versioned.
Every memory change is tracked. Roll back, diff, search — exactly like you do with code.
$ memgit log◉ c3a891f "Added React hooks rules + fixed routing memory" 2h ago│ memories: +3 added, 1 modified│◉ b7f3e2a "Removed stale API endpoint, updated auth context" yesterday│ memories: 1 removed, 2 modified│◉ a1d9f3c "Fixed prefer-callbacks rule — was causing errors" 2 days ago│ memories: 1 modified│◉ 9d1c4e8 "Initial import from claude-code (108 memories)" Jul 1memories: 108 added
Token Savings
95% fewer tokens. Same context quality.
claude.md loads everything on every session. memgit uses BM25 relevance scoring to load only the 8 memories most relevant to your current task.
* Numbers from memgit stats on a real 108-memory store
Multi-IDE Support
One memory store. Works everywhere.
Write a memory in Claude Code. Your Cursor session has it in 2 seconds. Your Gemini session too.
Auto-sync hook on session end. 5 native tools: search, write, forget, log, diff.
memgit injects relevant context into .cursorrules before each session.
TreeView sidebar, memory diff viewer, vscode.lm tool registration.
Same injection pattern as Cursor. Auto-detects Windsurf config path.
Section-marker injection + native tool definitions for Gemini 2.5 Pro.
AIAssistantContextProvider + Tool Window (Active / History / Threads).
Injects relevant context into .openai/instructions.md before each run.
OpenAPI 3.1 HTTP server. Any LLM with tool-use can call memgit's REST API.
The Future of Git
Git was invented because code needed history.
memgit was invented because AI context does too.
In 1991, Linus Torvalds was frustrated. Code was shared as patches emailed between developers. No way to see who changed what, no way to roll back a mistake, no way to merge two people's work. So he built git — every change became a permanent, addressable object.
That same frustration exists today — but for AI context.
Your AI assistant accumulates knowledge: your architecture decisions, your team's coding conventions, the bugs you've already solved. But when the session ends, it's gone. The next session starts blank. Your colleague's Cursor has no idea what your Claude figured out last week. When a “helpful” instruction turns out to be wrong, you can't roll it back.
We're in 2005. The year before git. For AI context management.
memgit is the missing layer. Every piece of AI context becomes a versioned object with a hash. Every session creates a checkpoint. memgit diff HEAD~1 HEAD shows what your AI learned today. memgit push syncs your entire context to your team.
The teams that figure this out first will have an AI advantage that compounds with every session. Their AI knows what yours doesn't. That gap widens every week.
In ten years, every engineering team will have a shared AI memory repository — versioned, audited, and as foundational as their code repository. memgit is the tool that makes that possible today.
Security & Privacy
Built to be trusted.
Your AI context contains your architecture decisions, your team's conventions, your business logic. We take that seriously.
Local-first. Always.
Your memories live in .memgit/ on your own machine. By default, zero bytes leave your computer. The free tier is completely offline — no account, no server, no network calls.
Verified: no outbound connections in free tier100% open source
Every line of the CLI, TOON parser, MCP server, and VS Code extension is MIT licensed and on GitHub. Read the code, audit it, fork it. Nothing is hidden behind a build step.
github.com/code4161/memgitHuman-readable format
TOON is plain text. Open .memgit/objects/ in any editor. grep it, diff it with standard tools, commit it to your own git repo. No binary blobs, no proprietary encoding.
$ cat .memgit/objects/ab/c123 → readableZero telemetry
The CLI sends no analytics, crash reports, or usage metrics. There is no call-home mechanism. You can run memgit on an air-gapped machine and it works identically.
Confirmed: grep -r 'telemetry' — returns nothingNo vendor lock-in
Export your entire memory store to markdown anytime with memgit export. Your context was yours before memgit, and it stays yours. Cancel Sync and your local data is untouched.
$ memgit export --format=markdown → your data backVerified integrity
Every memory object is SHA-256 content-addressed. memgit fsck verifies the entire store locally. The same cryptographic model git uses — no silent corruption possible.
$ memgit fsck → OK (108 objects verified)About cloud sync (coming Phase 4)
When Sync launches, your memory objects will be transmitted over TLS 1.3 and stored encrypted at rest using AES-256. You authenticate with OAuth (GitHub) — we never see your password. You control your data: delete your account and all objects are purged within 24 hours. The sync tier will publish a full security doc before launch.
Found a security issue? Report it privately on GitHub →
Get Started
Start in 30 seconds.
Free. Open source. No account needed. Works with your existing Claude Code memories.
1Install
2Set up your AI tool
memgit initInitialize and import your existing AI memory filesmemgit setup claude-codeWire MCP server + auto-commit hook (or: memgit setup all)memgit search "your query"AI now reads only the relevant memories — 95% fewer tokens