Skip to content

GuideUpdated September 2026

How to give Claude, Copilot, ChatGPT and Codex a persistent memory over MCP.

An MCP memory server is a local program that speaks the Model Context Protocol and exposes memory as tools an AI client can call: recall, checkpoint, search. Central Brain runs one per project on your machine. Connect it once and Claude Desktop, Claude Code, GitHub Copilot or the ChatGPT desktop app can read your decisions and files before they answer and save new notes as they work.

What the server exposes

Central Brain's MCP server is scoped to one project and exposes eight tools. The file tools are read-only; memory_checkpoint appends notes and memory_compact replaces the project brief. Which tools a client shows and calls varies by client.

central_brain_pingstatus

Confirms the server is running and reports the project root folder.

memory_recallmemory

Searches decisions, facts, open questions, todos and the project brief. Call it before answering anything about earlier work.

memory_checkpointmemory

Appends one note at a time (a decision, fact, question, todo or summary), so it is still there after the chat is compacted or closed.

memory_compactmemory

Replaces the project brief with an updated summary and marks older notes as compacted. It does not delete them.

semantic_searchsearch

Vector search by meaning over the project's prepared files, using the bundled local MiniLM model. Hits carry path, page and chunk where available.

search_textsearch

Exact text search across the files for when you know the phrase.

read_filefiles

Reads one file under the project root.

list_filesfiles

Lists files under the root, optionally in a subfolder.

Setup in five steps

  1. Install Central Brain and activate your license

    Download the installer from the location provided with your purchase, run it, and activate your license in the app.

  2. Create a project

    Choose the source folders and where prepared output should live. Central Brain prepares and indexes the files locally.

  3. Open the project's MCP setup

    Central Brain generates a project-specific MCP server entry for the client you pick: Claude Desktop, Claude Code, GitHub Copilot or the ChatGPT desktop app.

  4. Add the entry to your client and restart it

    Paste the generated entry into the client's MCP configuration, or let Central Brain write it, then restart the client so the tools appear.

  5. Run central_brain_ping

    Ask the assistant to call central_brain_ping. It confirms the connection and reports the project root. From now on, recall before answering and checkpoint after deciding.

Central Brain generates the exact server entry for your project, including the installed executable path and the project root, so you do not have to hand-type a path. The server runs as a local process over stdio with two arguments: --run-mcp-stdio and --root pointing at the project folder. For Claude Desktop on Windows the entry looks like this.

claude_desktop_config.json 
{
  "mcpServers": {
    "central-brain-meridian-renewal": {
      "command": "C:\\Program Files\\NeuroAIgent\\Central Brain\\NeuroAIgent.CentralBrain.exe",
      "args": [
        "--run-mcp-stdio",
        "--root",
        "C:\\Projects\\meridian-renewal"
      ]
    }
  }
}

Use one entry per project, each with its own --root. The install path varies by machine, and Central Brain writes the entry for you.

GitHub Copilot in VS Code reads a similar entry from its MCP configuration. In the ChatGPT desktop app, add the same command as a local (STDIO) MCP server in settings (OpenAI’s steps); it works in Chat, Work and Codex modes. In Claude Code, run claude mcp add central-brain -- followed by the same command. Capabilities vary by client. The getting started guide covers each client.

The recall and checkpoint loop

Tools only help if the assistant uses them. Central Brain's tool descriptions tell the model when to call each one, but whether a client follows them varies. To make it explicit, add a standing instruction to your client's system prompt, custom instructions or project file. Here is a short one.

Standing instruction 
At the start of every session call central_brain_ping, then memory_recall for the project brief and open items.
Before answering any question about earlier work, decisions or status, call memory_recall.
When we decide something, learn a fact, open a question or agree a todo, call memory_checkpoint.
When many notes have built up since the last brief, call memory_compact with an updated summary.
Prefer semantic_search over guessing what a file says. Quote the path and page.

The effect is a session that begins with "here is where we left off" instead of "how can I help you today". It survives compaction because the important parts were written down before the summary happened.

One project, many tools

Because the server is scoped to a project rather than to a client, you can connect Claude Desktop and Copilot to the same project and they share the same memory. A decision checkpointed in a Claude session is there for Copilot to recall an hour later. Keep separate projects for separate clients or matters, and each one keeps its own memory and index.

Troubleshooting

  • The tools do not appear. Restart the client after editing its configuration. Make sure Central Brain is running and the project is open.
  • ping says the index may be stale. Run Refresh in Central Brain. Only new and changed files are processed.
  • ping says "Not a Central Brain project". The server root points at a folder that has not been prepared yet. Build the project in the app first.
  • Recall returns nothing. The memory is empty until the first checkpoint. Ask the assistant to checkpoint the current state.
  • License verification. Activation and a license check every 14 days need internet access; in between, Central Brain runs air-gapped. Team and Enterprise plans can get a build with no online check. Cloud-hosted AI clients and optional cloud connections also need internet.

Still stuck? Central Brain support is a human.

The AI doesn't own the memory. You do.

Central Brain runs on Windows and macOS (Apple silicon). In the default workflow, files, prepared output, vectors and memory stay on storage you choose.

Get Central Brain Read the setup guide