Recall

A repository, once initialized, never has to explain itself again.

Recall initializes a repository's memory for AI coding assistants. Run one command, and every assistant — Claude Code, Codex, Cursor, Gemini CLI, ChatGPT — starts each session already knowing your stack, structure, commands, and conventions, instead of asking you again.

Quick start

npx recall-context init

Scanning repository...
Detecting architecture...
Generating AI memory...
Done.

That's it. The repository is initialized.

The npm package is recall-context; it installs the recall and recall-context commands, which invoke the same CLI. After npm install -g recall-context, the same thing is recall init.

Requirements

• Node.js 22 or newer.

What it creates

.recall/
├── memory.json        # Project Memory: the single source of truth
└── agents/
    ├── claude.md      # Claude Code
    ├── codex.md       # OpenAI Codex (AGENTS.md style)
    ├── cursor.md      # Cursor (project-rule front matter)
    ├── gemini.md      # Gemini CLI
    └── openai.md      # ChatGPT / OpenAI API

Nothing else, anywhere. memory.json holds what Recall found — metadata, technology stack, architecture, structure, commands, conventions, dependencies, and risks — and every entry cites the file that proves it:

{
  "name": "test",
  "command": "vitest run",
  "evidence": [{ "path": "package.json", "reason": "scripts.test declared as \"vitest run\"" }]
}

Each file in agents/ is a mechanical rendering of memory.json in the format that assistant reads best. No agent file contains anything memory.json doesn't.

Use it with your assistant

Assistant	How it picks up its file
Claude Code	Add @.recall/agents/claude.md to your CLAUDE.md
Gemini CLI	Add @.recall/agents/gemini.md to your GEMINI.md
Codex	Copy .recall/agents/codex.md to AGENTS.md (or reference it from there)
Cursor	Copy .recall/agents/cursor.md to .cursor/rules/recall.mdc — it's always applied
ChatGPT / OpenAI	Paste .recall/agents/openai.md into a Project's instructions or a system prompt

Recall complements a hand-written CLAUDE.md or AGENTS.md, it doesn't replace it: keep writing the things only a person can decide there, and let Recall supply what can be read straight off the repository.

Guarantees

• No network access. Everything happens offline, on your machine.
• No code execution. Recall reads files; it never runs your code, tests, or scripts.
• Writes only .recall/. No other file in your repository is touched.
• All or nothing. Either the complete .recall/ is written, or nothing is.
• Deterministic. The same repository state produces byte-identical output, on any machine.
• Idempotent. Running again on an unchanged repository changes nothing.
• No account, no API key, no LLM calls, and the CLI sends no telemetry. The documentation website is a separate application with its own operational analytics.

recall init refuses to overwrite an existing .recall/ (exit code 4, nothing changed). Pass --force to regenerate it after your project changes.

What it understands today

• Languages: TypeScript, JavaScript, Python, Go, Rust
• Frameworks: Next.js, NestJS, React, Express, Fastify, Vue, Angular, Nuxt, Svelte, SvelteKit, Astro, Remix, React Native, Electron, Hono, Koa
• Package managers: pnpm, npm, Yarn, Bun, pip, Poetry, uv, Pipenv, Go modules, Cargo
• Layouts: single packages and monorepos (workspaces, pnpm-workspace.yaml)
• Tooling: lifecycle scripts, Vitest/Jest/Mocha/AVA/Jasmine/Playwright/Cypress, ESLint/Prettier/Biome
• Risks: missing CI, a declared test framework with no tests, no license, no lockfile, unignored .env files

When the evidence for something is absent, Recall says nothing about it rather than guessing.

Learn more

• CLI reference — the command, the flag, and every exit code
• Memory format — memory.json and the agent files, field by field
• Architecture — how Recall is built
• Product specification — what Recall is, guarantees, and will never do
• Example output — Recall's own .recall/, generated from this repository
• Changelog · Contributing · Security

License

MIT