agent-memory-connect
v0.1.4
Published
Connect any MCP agent to Portable Agent Memory. Set up once with a blind-typed passphrase stored in the OS keychain; agents then talk to a local proxy that attaches the identity headers, so no secret ever sits in a config file.
Maintainers
Readme
agent-memory-connect
The friendly door to Portable Agent Memory. Set up once per device, then wire any MCP agent with one line and no secrets in any config file.
npm install -g agent-memory-connect
agent-memory setupInstall it globally rather than reaching for npx. npx re-resolves the package from the registry on every launch, measured at 16.5s against 4.0s for the installed binary. Several MCP clients cap server startup at 30 seconds and drop the connection past it, which looks like the memory being broken rather than a slow launcher.
Or from a checkout of the repo:
npm install
node bin.mjs setupWhat setup asks
Three things, and only the first two need thought:
- Identity. The address that names your memory, such as an email. Not a secret; the same identity reaches the same memory from every device and every agent.
- Passphrase. The only key to your notes. Typed blind (nothing shows while you type), confirmed twice, and stored in the operating system's credential store: Credential Manager on Windows, Keychain on macOS, the Secret Service on desktop Linux. It never sits in a config file, an env var, or a shell history.
- Endpoint. Just press Enter. The default is the hosted service at
https://agentsqa.xyz/mcp; you only type something here if you run your own deployment.
Setup then verifies the passphrase against your existing notes on the spot, so a typo surfaces immediately instead of masquerading as an empty memory. A brand new identity simply reports that no notes exist yet.
On a headless server with no credential store, setup says so and offers the honest fallback: a file at ~/.agent-memory/secret readable only by your account (permissions 600), written only with your explicit yes. status always tells you which store holds the passphrase.
Then add the printed line to any agent, for example:
claude mcp add -s user agent-memory -- agent-memory-s user makes the memory available in every project on the device; drop it to wire only the current project directory.
For Codex, add a local server to ~/.codex/config.toml:
[mcp_servers.agent-memory]
command = "agent-memory"
startup_timeout_sec = 60Codex is strict about its 30 second startup budget and reports MCP client for agent-memory timed out when a launch overruns it. Installing globally is the fix; the raised timeout is belt and braces.
Any other MCP client that launches local servers uses the generic form: { "command": "agent-memory" }.
On Windows, a client that cannot launch a .cmd shim directly needs { "command": "cmd", "args": ["/c", "agent-memory"] }.
The agent talks to a local MCP proxy over stdio. The proxy reads the passphrase from the credential store at runtime and attaches the identity headers to every HTTPS request to the memory endpoint. Every agent on the device shares the same memory; none of them ever sees the passphrase.
Other commands:
agent-memory statusshows the configured identity and endpoint, and whether a passphrase is held. Never the value.agent-memory resetremoves the stored identity and passphrase from the device.
Self-hosters pass their own endpoint during setup; everything else is identical.
