Deploy Mnemosyne

Mnemosyne runs as a scoped local library, an MCP service or a local HTTP service. Use a reviewed source checkout or an explicitly selected package release; website updates do not publish a new package version. The local runtime requires Node 22.16 or later. Keep persistent memory outside the checkout and use an explicit workspace and agent identity for each caller.

Build and run locally

From a source checkout, build the CLI and point it at your chosen database. This example starts an MCP process over its standard streams; the client owns that connection. It does not expose a network port.

npm ci --ignore-scripts
npm run build
node dist/cli/index.js mcp --db ./memory.sqlite \
  --workspace my-project --agent assistant

Scope and persistence

Workspace and agent selectors come from trusted host configuration. They do not authenticate an end user. The HTTP service defaults to loopback with bearer authentication; remote authentication and protected transport remain deployment responsibilities. Container deployments must retain their data volume and keep the memory directory writable only by the intended operator.

Existing installations

The local engine does not automatically import, rename or delete an existing backend. Keep existing collection identities and storage volumes intact while evaluating a separate local database. Use gradual migration for coexistence or full migration for a reviewed bulk move. Existing backend integrations retain their actual runtime identifiers and configuration contracts.

Export, recovery, and forgetting

A scoped export is different from a whole-database backup. Inspect omitted-record counts when exporting, and use recovery operations for a consistent complete snapshot. Backups contain plaintext memory content and require protected storage. Forgetting affects the live store; prior exports, backups and context already delivered elsewhere remain separate copies.