Give Claude Code a memory between sessions that you can see

You run /clear, or the terminal closes, or you come back after the weekend. The new Claude Code session is smart, but it has no idea that yesterday you stopped halfway through a refactor with two steps left and one thing to check by hand.

There are many fixes for this: handoff files, memory files, skills that write a summary. They all have the same weak spot. The memory sits in a file that you do not look at. You forget it exists, and so does the next session unless you tell it.

This guide shows a different place for that memory: a to-do list that stays on top of your screen, that you edit, and that the agent reads and writes over MCP.

In short

• Turn on the Focus Pad MCP server in DevWorkspace Pro and connect Claude Code with one copied command.
• Add two rules to your CLAUDE.md: read the project's collection at the start, write open steps at the end.
• The next session calls list_items, sees what is left, and continues.
• You see the same list all day, so the memory is never hidden.

Why a visible list works better than a hidden file

A handoff file is written for the agent. A to-do list is written for both of you.

When you complete a task yourself, it leaves the list. The next session does not repeat work you already did, because the task is gone.

Set it up

1. Turn on the server

1. Open Preferences → Focus Pad.
2. In MCP server for AI agents, turn on the switch.
3. Click Save.

The card shows Running on 127.0.0.1:47821. You do not need to restart the app.

2. Connect Claude Code

Click the copy icon next to Connect Claude Code and run the command in a terminal:

claude mcp add --transport http --scope user dwp-focuspad http://127.0.0.1:47821/mcp --header "Authorization: Bearer <token>"

The app fills in your token and port. --scope user makes the server available in every project, so you do this once. In Claude Code, /mcp should list dwp-focuspad as connected.

3. Make a collection for the project

In Focus Pad, press ⌘ + K, type the project name and press Enter. Or ask the agent: "Create a Focus Pad collection called billing-api." It calls create_collection.

4. Add the rules to CLAUDE.md

## Focus Pad

- At the start of a session, read the Focus Pad collection "billing-api" with list_items.
  Treat its tasks as the open work for this project. Ask me which one to start with.
- When you finish a task from that list, complete it with delete_item.
- Before I end the session, or when I say "wrap up", add every open step and every
  manual check to that collection with add_item. One task per step, in order.
- Use priority 5 only for steps that block a release.

The tool descriptions already tell the agent the formats for priority, color, deadline and Pesky, so you do not need to explain them.

What a session looks like

Evening, end of session. You type "wrap up". The agent adds three tasks:

• Finish moving invoice totals to the Money value object (2 of 5 files left)
• Run the full test suite, InvoiceTotalsTest was flaky once
• Check the staging webhook log by hand after deploy

Next morning, new session. You open Claude Code in the project. Because of the rule, it calls list_items first and replies:

There are 3 open tasks in billing-api. The first is to finish moving invoice totals to Money, with 2 of 5 files left. Start there?

You say yes. It works, then completes the task with delete_item, and the task leaves your window.

During the day. You notice the flaky test is fixed upstream. You Cmd+click the task to complete it. The agent does not see this until it reads the list again, which it does at the next "wrap up" or the next session.

Tips that make it work better

• Write tasks the agent can act on. "Finish Money migration in InvoiceService and CreditNote" is better than "finish refactor". The agent cannot add notes under a task, so the one line must carry the detail.
• Add a deadline to manual checks. Ask: "Add the staging webhook check, due today 16:00." The task label turns amber when 24 hours or less are left and red when the deadline is past.
• Give tasks different text. To change a task, the agent sends its position and its current text. If two tasks have the same text and swap places, the check cannot tell them apart.
• Do not type in the same collection while the agent writes. If you both change the list at the same moment, one change can be lost. Wait for the agent to finish.
• Use one collection per project. The agent reads one collection at a time. Separate lists keep each session focused. See per-project to-do lists.

What this does not do

Be clear about the limits, so the list stays trusted:

• The agent works only with tasks. It cannot change your headings or notes, or rename or delete collections.
• It does not record why it made a choice. If you need decisions and failed attempts, keep those in your repo docs and use Focus Pad for what is left to do.
• DevWorkspace Pro must be open. If it is closed, the agent gets a connection error.
• Completed tasks are removed, not archived. The list shows open work only.

Other MCP clients

The server speaks Streamable HTTP and needs an Authorization header. Any client that supports both can connect with the same rules in its own instructions file. The setup values are in the Focus Pad MCP docs.

For the full tool list and safety details, see Give your AI coding agent a shared to-do list. To pick up the conversation itself after a restart, see resume Claude Code sessions.