Give Claude a memory that survives your sessions

Go to notebook.google.com, sign in, create a notebook and name it exactly Claude Memory. Leave it empty: the skill is what fills it.

The name matters. The routing rule in step 4 will point at it character for character. Pick it now and leave it alone.

The bridge between the two is an MCP server. An MCP is a connector: it gives Claude a capability it does not have out of the box, here talking to your notebook.

Two commands, in your terminal. The first opens a Chrome window where you sign in to Google, once:

npx -y -p @roomi-fields/notebooklm-mcp notebooklm-mcp-setup-auth

The second registers the server in Claude Code:

claude mcp add notebooklm -- npx -y -p @roomi-fields/notebooklm-mcp notebooklm-mcp

Quit and reopen Claude Code, then ask it to list your notebooks. If it returns your list, you are wired.

This is the piece that makes the memory automatic. Without it nothing gets saved and you just have an empty notebook wired to Claude.

Create the file .claude/skills/wrap-up/SKILL.md in your project, and paste this in:

---
name: wrap-up
description: Saves the current session into the Claude Memory notebook. Trigger when the user says "wrap up", "save the session", "let's stop here".
---

# Wrap up

When invoked, write a summary of the current session and push it
as a new source into the "Claude Memory" notebook.

## What the summary contains

- Today's date and the project concerned.
- What was decided, with the reason behind each decision.
- What was done, file paths included.
- What is still open, and the next action.
- Dead ends: what was tried and did not work.

## Format

A title "Session DD/MM/YYYY, <project>", then the five sections
above as short lists. No transcript, no pasted code: a summary
readable in thirty seconds six months from now.

## Writing

Add the summary to the "Claude Memory" notebook through the
notebooklm server's add-source tool, as text. Confirm in one line,
with the title of the source created.

If the notebook cannot be found, list the available notebooks and
ask which one to use, never create one more.

The part that makes the difference is the "Dead ends" section. That is what you lose systematically between two sessions, and it is what makes you retry the same thing three weeks later.

Saving is not enough, Claude still has to read. Add this block at the very top of your CLAUDE.md:

# Memory

At the start of every session, query the "Claude Memory" notebook
in Gemini Notebook before asking me to re-explain anything. Ask a
precise question about the current project and pull back only the
answer, never the full sources.

At the end of a useful session, offer the wrap up if I have not
asked for it.

The placement is not cosmetic. A CLAUDE.md is a rules file loaded at startup, and the higher an instruction sits, the more reliably it is followed. Buried at the bottom, the rule gets skipped half the time.

The second sentence of the block is the one people leave out: without it, you have to remember to say "wrap up". With it, Claude offers it to you.

Your memory notebook does double duty. Everything the skill saved into it becomes raw material, and Gemini Notebook knows how to turn it into something else for free.

The use case worth the detour if you publish: you work on a project for two weeks, every session gets saved, and at the end you ask for an infographic or a video overview. Your content feeds on your actual work, with no extra production.

For a freelancer, the same mechanic pulls a quarter-end report out of your client follow-up sessions, with nothing to retype.

Five boxes to tick once the install is done:

If Claude cannot see the notebooklm tools, the server did not register: run the add command again and fully restart Claude Code. If the wrap-up writes nothing, it is almost always the notebook name not matching the one in the routing rule exactly. And if the answers feel stale, wrap up more often: the notebook only knows what you put in it.