@ndestates/orchestrator
v2.3.10
Published
Stop re-pasting your entire repo. Local AI project memory + token cache for Claude, Grok, Cursor, Copilot & more. One brief → every session. Free. Private. Apache-2.0.
Maintainers
Readme
Orchestrator Memory — AI Token Cache for VS Code & Cursor
Local project memory + AI token cache.
Stop re-pasting your entire repo into Claude, Grok, Cursor, Copilot, Gemini or ChatGPT.
One short, reusable brief → every session. Fewer tokens. Faster start. Real continuity. Everything stays on your machine.
Free · Apache-2.0 · 100% local · Bring Your Own Keys (BYOK)
Install (only method): npm install -g @ndestates/orchestrator
https://www.npmjs.com/package/@ndestates/orchestrator
Why AI coding assistants forget (and how this fixes it)
Every new chat starts from zero.
You re-explain architecture, current branch, decisions, “don’t touch X” rules, and project conventions — burning tokens and killing momentum.
Orchestrator Memory is a local AI memory layer and token cache for coding. It stores durable project context in a small SQLite database plus an optional vault dual-write so any assistant (Claude, Grok, Cursor, Copilot, Gemini, ChatGPT, and others) can start from a lean, high-signal brief instead of a wall of context.
Result:
- Dramatically lower token use
- Session start in seconds instead of 10 minutes of pasting
- Better continuity across tools and days
- Privacy by design — no cloud upload of your memory or vault data
Core features
| Feature | What it does |
|---------|--------------|
| SQLite runtime memory | Fast project (or global) token cache at reports/memory/memory.db (gitignored) |
| CONTEXT.md | Git-tracked lean snapshot so memory survives clone and branch switch |
| Vault dual-write | Durable, scrubbed lesson graph in reports/vault/events.jsonl |
| Slash commands everywhere | /session-start, /memory-brief, /chain, /quick-start, /setup work in Command Palette, Chat (@orchestrator), and any model |
| One-click brief | Generates a compact brief, copies it, and writes SESSION_READY.md + MEMORY.md |
| Lazy activation | Zero work on window open — activates only on command, keybinding or Chat |
| Inbox ingest | Drop files into reports/memory/inbox/ for automatic capture |
| Query / Seed / Status / Serve | Ask the store, snapshot the situation, inspect health, or run a local HTTP endpoint |
| Copy targets | One-click for Claude, Cursor, Copilot or export as MEMORY.md |
| 100% local & private | Memory and vault never leave your machine |
Wedge: Not another AI chat panel. Portable project memory that cuts tokens across every assistant you use.
Quick start (under 60 seconds)
- Install Node.js 18+ and Python 3.10+ if needed.
- Install once:
npm install -g @ndestates/orchestrator
orchestrator memory brief --seed- Optional — write template files in one app:
npx orchestrator init . --no-pr- Paste the brief into Claude, Grok, Cursor, Copilot, Gemini or ChatGPT.
Fleet/wave scripts are deleted. Do not use uv, pip, or VSIX to install.
How storage works
ingest / seed / brief → reports/memory/memory.db (SQLite — fast token cache)
↘ dual-write → CONTEXT.md (git-tracked snapshot)
↘ dual-write → reports/vault/events.jsonl (vault — durable lessons)- SQLite = runtime memory used for briefs and queries (project-scoped or global).
- CONTEXT.md = committed lean snapshot; restored across branches by session-start persist heal.
- Vault = append-only, scrubbed event graph for long-term lessons.
- Both stay local. Extension never uploads memory or vault content.
Full details live in the built-in INSTRUCTIONS.md (Command Palette → Orchestrator: Read instructions).
Commands cheat sheet
| Action | How |
|--------|-----|
| All slash commands | Command Palette → type / or Ctrl/Cmd+Shift+O |
| Session start (brief + chain) | /session-start or Ctrl/Cmd+Shift+S |
| Lean brief only | /memory-brief or Ctrl/Cmd+Shift+M |
| Install host CLI | /setup or Orchestrator: Setup / Install CLI |
| Quick start (CLI + seed + brief) | /quick-start |
| Help | /orchestrator-help |
| Chat | @orchestrator /session-start |
After a brief you get the text on the clipboard plus optional SESSION_READY.md and MEMORY.md in the workspace.
Requirements & privacy
- Host CLI: npx (preferred) or
orchestratoron PATH. VSIX/Marketplace is parked. - BYOK — you use your own Claude, Grok, OpenAI, Gemini, Copilot or Cursor accounts. We do not provide, sell or proxy API keys.
- Works with any AI that accepts pasted text.
- Memory and vault data remain on your machine only.
Keywords & discoverability
AI memory · token cache · project memory · local AI · Claude memory · Cursor memory · Copilot context · Grok coding · reduce tokens · session continuity · SQLite memory · durable context · VS Code AI extension · Cursor AI extension
License
Apache License 2.0 — free and open source.
Stop re-prompting the whole repo.
Install Orchestrator Memory, run /session-start once, and give every AI the same high-signal project brief.
