Databases

SQLite by default, with no setup and no network. The connection string picks the backend, so moving to Postgres or Mongo is one command.

SQLite

Stored at ~/.claude-memory/memory.db. It comes from Node's builtin node:sqlite, so there is nothing to install and nothing compiles. Memory is filed by project, so one database serves every repo you work in.

Switch database

claude-db use "postgres://user:pass@host:5432/memory"
claude-db use "mongodb+srv://user:pass@cluster.mongodb.net/memory"

use switches the backend and verifies it can be reached before saving. Install the driver you need first, either npm install -g pg or npm install -g mongodb, since neither ships by default and claude-db looks for them next to itself. CLAUDE_DB_URL overrides the config file when you need a one-off.

Sync two databases

claude-db sync "postgres://user:pass@host:5432/memory"

A two-way merge, so a laptop and a desktop can each keep working offline and reconcile later. export and import cover the same ground as files when you would rather move a dump by hand.

Teams and several machines

Point everyone at one Postgres or MongoDB database and they share memory.

• Projects are matched by git remote. A repository with an origin remote is one project wherever it is cloned, so a laptop and a desktop share memory. The ssh and https addresses of one repository match, and no token in the address is ever stored. A folder with no remote is matched by its path. A fork has its own remote, so it is its own project. Set project.remote to use a remote other than origin.
• Update every machine to 0.11.0 or newer. An older version does not see notes shared by git remote.
• Personal rules stay personal. A rule about how you like to work is filed under your git email and never reaches a teammate. Set git config user.email so it is filed under the right name.
• Use a restricted account. The account should reach only its own database.

Memory saved before 0.11.0 under a folder path is still found and is not moved.

MongoDB Atlas vector index

Atlas can index vector search instead of scanning brute-force, but Atlas indexes are not created by this tool. Add one by hand on the observations collection:

{
  "name": "memory_vector",
  "type": "vectorSearch",
  "fields": [
    { "type": "vector", "path": "embedding", "numDimensions": 256, "similarity": "cosine" },
    { "type": "filter", "path": "project" },
    { "type": "filter", "path": "kind" },
    { "type": "filter", "path": "tags" },
    { "type": "filter", "path": "createdAt" }
  ]
}

numDimensions is 256 for the builtin embedder and 384 once @xenova/transformers is installed. Without this index Mongo still works: vector search falls back to scoring candidates in process.