Kemory

kemory

The command-line client for Kemory, a persistent memory vault for AI agents. It signs you in, connects MCP clients such as Claude Code, Codex, Cursor and VS Code to your vault, and runs the MCP bridge they use.

Quick start

You need a Kemory account; sign up at kemory.sekondbrain.ai.

1. Connect your AI clients. Nothing to install first:

    npx -y kemory connect   # Node.js 18+
    uvx kemory connect   # or with uv

   Your browser opens to sign in. kemory then lists the AI clients on this machine, [ready] or [connected], and asks once: connect every ready client, choose some, or cancel. Claude Code and Grok Build get the Kemory plugin; Codex, Copilot CLI, Cursor, Gemini CLI, OpenCode and VS Code an MCP entry plus automatic memory; the others an MCP entry. Then the connection is checked. A summary lists what is left to do, such as which clients to restart.

2. Restart the clients it lists. Warp picks the change up without a restart.
3. Check the setup end to end at any time with kemory doctor.

--dry-run shows what would change without changing anything. In scripts, name the clients instead: --host all, or --host claude-code, --host codex and so on (MCP clients lists them); kemory then asks nothing.

Install

brew install sekondbrainailabs/s9n/kemory
npm install -g kemory
uv tool install kemory

Builds exist for macOS (Apple silicon and Intel), Linux (arm64 and x64) and Windows (x64). They are attached to each GitHub release with checksums and build provenance (gh attestation verify <file> --repo SeKondBrainAILabs/kemory-cli). kemory upgrade tells you whether a newer release exists and how to update the copy you installed.

Commands

Every command accepts --output json for scripts and agents (stdout then carries only the JSON payload), --quiet, --no-color and --env <name>.

MCP clients

kemory connect --host <name> and kemory mcp install --host <name> connect one client, and --host all connects every client installed on this machine.

For Claude Code and Grok Build they install the Kemory plugin through the client's own claude or grok command. Its hooks save each session to Kemory, and in Claude Code also load your memory when a session starts. Without that command, or for an environment other than prod, they write the MCP entry instead.

For Codex they also add hooks to ~/.codex/hooks.json that load your memory when a session starts and save the session as it goes (Hooks). Codex runs a new hook only after you approve it, so open Codex once and approve them with /hooks. Your other hooks in that file are left as they are.

For Cursor they add the same hooks to ~/.cursor/hooks.json, and for Gemini CLI to the hooks in ~/.gemini/settings.json. Both run them without asking, and your other hooks are left as they are.

For VS Code and the Copilot CLI they write ~/.copilot/hooks/kemory.json, which both run (in VS Code through the chat.useHooks setting, on by default). kemory owns that file and rewrites it on every connect.

For OpenCode they add the opencode-kemory plugin to the plugin list in opencode.json, which OpenCode installs from npm. It does the same: it adds your memory to the system prompt when a session starts and saves your prompts as the session goes. Connected to an environment other than the default, they write the plugin as ~/.config/opencode/plugins/kemory.js instead, which connecting again replaces, so keep your own plugins in other files.

For every other client they write an MCP entry. The entry only starts kemory mcp serve; no token or key is written into the client's config, and everything else in the file is left as it was.

On Windows the files under Library/Application Support and .config live in %APPDATA% instead, except OpenCode's, which stay in ~\.config\opencode. Grok also reads ~/.claude.json, so the claude-code entry reaches it too.

One connection per client

Each client should reach Kemory one way. Two ways give it every Kemory tool twice, which costs context in every session.

connect and mcp install write nothing for a client that already has Kemory: Claude Code or Grok Build with the Kemory plugin enabled, or Claude Desktop when your account has a claude.ai connector in use. kemory doctor reports a kemory entry that duplicates either one, and kemory doctor --fix removes it.

Claude Desktop rewrites its config file from memory while it runs, which drops an entry added or removed in the meantime. When its config has to change while it is open, connect, mcp install and doctor --fix offer to quit it, make the change and open it again (doctor --fix --yes does so without asking). Run from the terminal inside the Claude app, which would close with it, they offer to open Terminal and continue there instead (macOS).

Copilot cloud agent

GitHub Copilot's cloud agent works in a sandbox where no one signs in, so it reaches Kemory with an API key. Run kemory connect copilot-cloud in the repository it works on, with gh signed in. It:

1. writes .github/hooks/kemory.json, the hooks that load your memory when a session starts and save the session (commit and push it);
2. creates an API key for the repository, or a new one on a re-run, and stores it as the repository's Agents secrets KEMORY_API_KEY (for the hooks) and COPILOT_MCP_KEMORY_API_KEY (for the MCP server);
3. prints the MCP configuration to paste in the repository's Settings → Copilot → MCP servers, and the Kemory host to add to Settings → Copilot → Internet access, so the hooks can reach it.

Without gh it prints the key once, to add as both secrets yourself. The hooks also take KEMORY_API_KEY anywhere else no one signs in.

Exit codes

Scripts and agents can act on why a command failed:

Configuration

Sign-ins are stored per environment in ~/.kemory/credentials-<env>. Extra environments can be defined in ~/.kemory/environments.json (or the file named by KEMORY_ENVIRONMENTS_FILE):

{"myenv": {"issuer": "https://idp.example.com/realms/x", "kemory_url": "https://kemory.example.com", "dashboard_url": "https://app.example.com", "server_name": "kemory-myenv"}}

Hooks

Coding agents with hooks can give a session your Kemory context when it starts and save it when it ends. kemory connect sets this up where it can; these are the commands the hooks call.

Neither can hold up the agent: any error is silent, the exit status is always 0, and each gives up after KEMORY_HOOK_TIMEOUT seconds (default 2). KEMORY_HOOK_DEBUG=1 prints why a hook did nothing.

Capture sends only your own prompts, never the agent's replies or tool output: at most the last 12 new prompts and 8000 characters per memory, with credential-shaped text (API keys, tokens, passwords, private keys) replaced by [REDACTED]. They are stored in the user:sessions namespace, visible only to you. It records what it has stored in ~/.kemory/.captured/<session>, the same file the Kemory plugin uses, so a session is never stored twice.

What the CLI sends

Commands talk to the selected environment's Kemory API, and login and token refreshes to its sign-in server (accounts.sekondbrain.ai for prod). Each API request carries your access token, the active organisation, the CLI version, OS and architecture (X-Kemory-Client), and the name of the command being run (X-Kemory-Command, never its arguments). Kemory uses the last two for product analytics; kemory telemetry off turns that off for your account. kemory doctor and kemory upgrade also read the public release list on GitHub to compare versions.

How SeKondBrain handles your account and vault data is set out in the Privacy Policy.

Uninstall

1. Remove the kemory entries from your MCP clients' config files (the MCP clients table names each file); for Continue, delete ~/.continue/mcpServers/kemory.json.
2. Run kemory logout, or delete ~/.kemory/ to drop every saved sign-in.
3. Uninstall the CLI the way you installed it: brew uninstall kemory, npm uninstall -g kemory or uv tool uninstall kemory.

Building from source

make build   # produces ./kemory
make check   # what CI runs: tests, lint, vulnerability scan, hygiene check

See CONTRIBUTING.md for how to work on the CLI and SECURITY.md for how it handles credentials and how to report a vulnerability. Questions and bugs go to GitHub issues.

License

MIT