RepositoriesCRBRO

MCP server · open source · MIT

CRBRO

Memory for your AI that survives the chat closing.

A local MCP server that remembers what you decided, how things are deployed and what is still pending, across sessions and across clients. It all lives as plain JSON on your own disk, and one command installs it in Claude Code, Claude Desktop, Codex and Cursor.

npm

  • 15 tools
  • Node.js 20+
  • Claude Code · Claude Desktop · Codex · Cursor
  • No account, no telemetry
The tour from the README: memory across sessions, corrections that replace old facts, stale values flagged, secrets kept out and the browser editor.

What it does

A memory you can read, correct and trust

What makes it more than a notes file the model reads at the start.

Corrections replace old facts

When something changes, the new fact supersedes the old one instead of sitting next to it. The model never has to pick between two versions of the truth.

Old values come with a warning

Versions, prices, ports, URLs: anything that can change carries a shelf life. Past it, recall sets the value apart as possibly_stale and says where to check it.

An honest “nothing stored”

A question with nothing clearly stored behind it comes back flagged no_answer, not padded with the closest guess.

Secrets kept out

Credentials go to the operating system's keychain and the memory keeps only their name. A redaction filter catches the ones that slip into a note.

Readable and editable

npx crbro-memory view opens a local page (127.0.0.1, one-time token) to search the memory, edit or retire an entry, merge two topics and close open items.

Yours, on your disk

Plain JSON in ~/.crbro, with no account and no telemetry. Export it to Markdown or JSON, or import a CLAUDE.md you already have.

Measured

Numbers with a date and a script

Measured for version 2.11.0 on 14 November 2026, on Windows 11 with Node 24. Each figure has a script in benchmarks/ that you can run yourself.

83%recall@1, keyword engine alone

48 blind paraphrased queries. recall@3: 90%.

24/24credentials caught

Adversarial cases, with 0 false positives on 28 near-misses.

≈10kcharacters to boot

9,973 characters to load a real 563-neuron memory at the start of a session.

1,337tests passed

107 files on Windows 11 (2.12.0). CI runs the suite on Linux, Windows and macOS.

What the benchmarks deliberately do not claim (human productivity, “it knows you”, comparisons with other memory systems) is written down in benchmarks/LIMITS.md (in Spanish).

What it doesn't do

Where it stops

  • It doesn't sync through a cloud. Sharing a project goes through a private git repository you own, and only the projects you choose.
  • Installing it doesn't wake it up: the assistant has to call crbro_boot at the start. install wires that for you; check that it ran in a new conversation.
  • Semantic recall is optional and takes a one-time download of about 500 MB. Without it, search is keyword-based, with a synonym table in English and Spanish.
  • It isn't free in context: the 15 tool definitions take about 7,500 tokens in a client that loads every tool.

Install

One command

Claude Code, Claude Desktop, Codex, Cursor

npx crbro-memory install

Finds the MCP clients on this machine, registers CRBRO in each one and makes the memory load itself at the start of every conversation. It merges into your config files and keeps a timestamped backup of each. Then restart the client and open a new conversation.

Claude Desktop, without a terminal

Download the .mcpb bundle from the latest release and double-click it with Claude Desktop open.

Check what is installed where

npx crbro-memory doctor

Read-only: it repairs nothing and reads no credential values.

You need

  • Node.js 20 or later (the Claude Desktop bundle needs none)

The code, the issues and the history are on GitHub.

Star it if it helps; open an issue if it breaks.

View on GitHub