How to add long-term memory to Claude Code

1. Prepare one project before connecting Claude Code

Choose a focused project with a few documents whose contents you already know. Select the source folder and a separate prepared-output location in Central Brain™. Run preparation and indexing, review failures, then search locally for a distinctive phrase. Confirm the result opens the intended source. This separates file preparation problems from client connection problems.

Keep unrelated clients or engagements in separate projects. A server name such as central-brain-project is only a label; the actual project root determines the memory and files it exposes. Check that root every time you create or duplicate a configuration.

2. Add the generated local MCP configuration

Open the project MCP setup in Central Brain™ and select Claude Code. Use its generated command or configuration to carry the actual installed executable and project root into the client. Anthropic confirms the local stdio command form below. The executable and root shown here are placeholders: replace both from your app-generated configuration before running it.

The two Central Brain™ launcher arguments are --run-mcp-stdio and --root. Preserve the argument boundaries and quoting supplied by the app. Run the command from the relevant project directory. Use a project-specific server name when keeping multiple project connections.

claude mcp add --transport stdio central-brain-project -- "<ABSOLUTE_PATH_FROM_CENTRAL_BRAIN>" --run-mcp-stdio --root "<ABSOLUTE_PROJECT_ROOT>"
claude mcp list
claude mcp get central-brain-project

3. Verify the connection and save a first checkpoint

This fresh-session check makes Central Brain™ persistence visible: inspect the saved decision, then inspect the tool result that retrieves it. Record the result as a setup check for your project.

1. Open Claude Code and inspect the connection with /mcp. Review the actual tools and any client permission prompts.
2. Ask Claude to call central_brain_ping and report the project root. Stop and correct the configuration if it identifies another project.
3. Ask memory_recall for the current project brief and open items. An empty result can simply mean that nothing has been saved yet.
4. Provide a harmless test decision and ask Claude to save it with memory_checkpoint as a decision. Inspect the saved record.
5. Begin a fresh session and ask Claude to recall that decision through Central Brain™ before you supply its contents again.

4. Make recall and checkpointing part of the work

Keep a short instruction in the instruction mechanism your Claude Code version supports. Tell the assistant to verify the project, recall relevant prior decisions, search source files when a claim needs evidence, and checkpoint newly agreed decisions. A request to remember something is useful only when the tool call actually writes it.

Save a concise decision, verified fact, open question, todo or summary rather than an unreviewed transcript. When the record becomes long, review a proposed project brief before using memory_compact. That tool replaces the brief and marks older notes as compacted; it does not delete the original notes.

At session start, call central_brain_ping and verify the project root.
Call memory_recall before answering about earlier project work.
Use semantic_search or search_text for source evidence.
Checkpoint agreed decisions and open items; review important notes with me.

Keep your project connection and memory current

Compare the project root in the generated configuration with the root returned by ping. Then check whether a checkpoint was actually written. A chat closing, compacting or switching models does not automatically create Central Brain™ memory. If file results are stale, refresh the project in Central Brain™ and inspect preparation failures.

For a failed connection, check the executable still exists, quote paths containing spaces, preserve each launcher argument, and inspect /mcp status. Keep source access narrow. Local indexing does not make a cloud Claude session offline: the retrieved context is sent to the connected assistant for its answer.

Recover when a recalled decision is out of date

Persistent memory is useful only when the record remains accurate. If a requirement changes, inspect the original source, save a clear correction with the current document version, and review the project brief. Ask the next session to retrieve the revised decision and distinguish it from the superseded note.

Keep stable coding instructions in the client’s supported instruction mechanism. Keep changing project decisions and supporting evidence in the project record. Re-run a known-answer recall after changing the server configuration, moving a project, or restoring a backup; successful connection status alone does not show which facts the assistant can retrieve.