Sign in

One memory across AI tools.

Decisions and project facts, recalled from Claude, Codex, Cursor or any MCP tool. Nothing is shared until someone who didn't write it approves it.

  • Claude Code
  • claude.ai
  • Claude Desktop
  • Claude mobile
  • Codex
  • Cursor
  • Gemini CLI
  • VS Code

Each connects over MCP, the open protocol AI tools use to reach other services. The Claude Code plugin is in daily use against this server; the others connect through each tool's own MCP settings (what has not been tested live).

waitingnew memory

Staging database passwords live in Vault. Never put them in a .env file.

  1. Savedby Claude Code
  2. Checkedno secret, no finding
  3. Approvedwaits for someone who didn't write it
  4. Sharednot yet

Approve it, and other AI tools can recall it.

How a memory gets approved

How a memory gets approved
As text
  1. An agent, here crew/scout, saves a memory.
  2. The memory waits, held: only its author and its reviewers can read it. It is scanned. A leaked secret refuses the write: a private key, a known provider's API token, a user:password pair in a URL or after curl -u, or a password set in config (DB_PASSWORD=…), and so does a field over 1 MiB. Prompt-injection phrases, hidden Unicode and long base64 blobs do not refuse it: each is recorded as a finding on the held memory and shown to its reviewer.
  3. Someone who did not write it, nor what it builds on, approves it: a person, here eugene, or a reviewer agent. The key that wrote a memory never approves it.
  4. The memory is promoted, approved and shared: every agent with access to its workarea recalls it. A workarea is a named area of memory, such as a team or a project, and a key reads only the workareas it is granted.
  5. Two opt-in exceptions, each confined to one workarea, are described under workareas.

Nothing shared unapproved

An agent's memory waits until a person, or a reviewer agent, who didn't write it approves it.

Every memory shows its source

Who wrote it, when it was approved, whether a person's key or an agent's key approved it and, where given, the repository, commit and file it came from. An administrator can read which reviewer.

Answers say why they matched

Each hit names the words it matched, which shows why it came back.

Get started

The public server is by invitation: email contact@doublegate.io. Invited? Sign in with GitHub or Google and issue an agent key on the Keys page, /web/keys. Then:

  1. Install oml on Linux. The installer checks the download against the release's signed checksums. On a Mac, build it from source; on Windows, use PowerShell (install options).
  2. Store the key. oml login asks for it, checks it with the server and keeps it in a file only its owner can read.
  3. Set up Claude Code. oml setup claude-code installs the plugin with that key. Then run /reload-plugins in Claude Code, or restart it.
  4. claude.ai or the Claude app? On the Keys page, issue a connector key and copy its connect code. On claude.ai, open Customize > Connectors, add https://memory.doublegate.io/mcp and paste the code. A phone gets it too.

No invitation? Try a local server in about five minutes. claude.ai and phones cannot reach a server on a laptop; they need one on a public HTTPS address.

Memories stay on the chosen server. On the public server its operator can read them; its request logs hold no memory text and are kept 30 days (what is logged).

Linux
# install the CLI (Linux)
$ r=https://raw.githubusercontent.com/e8kor
$ u=$r/open-memory-lab/v0.7.3/scripts/install
$ curl -fsSL --proto '=https' --tlsv1.2 \
    "$u/install.sh" | sh

# store the key; paste it when asked
$ oml login https://memory.doublegate.io

# install the Claude Code plugin
$ oml setup claude-code
# then, in Claude Code: /reload-plugins