Memory & scheduling

How Cowork remembers, and how to schedule it

Cowork has four ways of remembering. They are easy to confuse because the words around them overlap. This page sorts them out in one sitting, then shows the one concrete example you need: how to use /schedule for a recurring morning task.

§1 — The four persistence layers

The four persistence layers

The simplest way to keep them straight is to ask, for each one: who writes it, when does Claude read it, and what is it good for.

Layer 01

Project memory

Cowork writes it. Invisible to you.

Automatic, scoped to this Cowork project, stored locally on your machine — not in Anthropic's cloud, not synced to your other projects. Cowork notices things across sessions and tucks them away in a private store.

Excluded by design: passwords, financial details, ID numbers, anything Cowork classifies as sensitive. View, edit, or reset it from Project settings > Memory.

Layer 02

INSTRUCTIONS

You write it. Once, on day one.

The text you paste into Cowork's Instructions field. Acts as a constitution — set the rules of your workspace and rarely touch it again.

The Folder page has the full template. Cap: 300 lines. Edit when the rules of your workspace change — not when this week's priorities change. That is what the next layer is for.

Layer 03

MEMORY.md

Claude writes it. You can edit it.

The running whiteboard. Claude appends a short, dated entry every time it learns something durable about you — your voice, your stakeholders, decisions you have made. One or two sentences per entry.

Cap: 150 lines. When the file gets close, the oldest third moves to ARCHIVE.md, which Claude only reads when you point at it.

Layer 04

context/

You write it. Three short files.

The three files inside context/: about-me.md, work-preferences.md, and current-focus.md. Claude reads them on demand — pointed at by the Routing Map in INSTRUCTIONS.md.

current-focus.md is the working surface — the file you tweak when your priorities shift. The other two are steadier. Caps: 1500 / 500 / 300 words.

Summary. Project memory is automatic — leave it alone. INSTRUCTIONS is the constitution — set once, edit rarely. MEMORY.md is the day-to-day whiteboard. The context/ files are the long-form profile, read on demand.

§2 — How MEMORY.md grows in practice

How `MEMORY.md` grows in practice

You will not write to MEMORY.md yourself most of the time. Claude does. After a session in which it learned something durable — that you spell “Trelleborg” without an umlaut, that your end-of-quarter is the 15th not the 31st, that the contractor you keep referring to as “the Italians” is actually a Milanese reseller called Adriatech — it appends one or two sentences.

A few sessions in, the file looks like this:

# MEMORY

## 2026-05-22
James prefers single-sentence summaries with the conclusion first.
Polaris proposal goes to Anders on Thursdays, never Fridays.

## 2026-05-26
"Adriatech" = the Milanese reseller. Contact is Sara Conte.
Avoid the word "leverage" in client-facing drafts.

## 2026-05-28
EU CBAM deadline for Trelleborg is 2026-07-15, not 2026-07-31.

(... entries continue, dated, two sentences max ...)

When the file reaches 150 lines, the next session begins with Claude moving the oldest third to ARCHIVE.md and confirming the move in chat. You can intervene — “keep the Adriatech note, archive everything before 1 May” — or let it run on autopilot.

Why archive instead of delete? Most lessons stop being load-bearing after a few weeks but are still useful to recover when context comes back around. Archiving keeps the running whiteboard sharp; deletion would lose the history. The 150-line cap is the part that does the real work.

§3 — Scheduling recurring tasks

Scheduling recurring tasks

Cowork can run a task at a fixed time and write the result back to your folder. You set it up once and it runs unattended — provided Claude Desktop is open and your machine is awake at that time. If the machine is asleep when a task is due, Cowork skips that run and triggers it again when the app reopens.

You create a scheduled task with the /schedule command inside the Cowork project. The wording is informal — you describe when and what, Cowork interprets it. Scheduled tasks appear in a dedicated Scheduled section in the Cowork sidebar so you can review or stop them later.

Example — the morning briefing

The single most useful scheduled task for a personal-assistant folder. It reads your current priorities, drafts a one-page plan for the day, and writes it into daily-notes/ before you sit down.

/schedule every weekday at 8:00am: read MEMORY.md and
context/current-focus.md, plus any files I have added or edited in
projects/ since yesterday. Write a one-page plan for today into
daily-notes/YYYY-MM-DD.md with this structure:

  ## Top three for today
  <your suggested ordering of the priorities in current-focus.md>

  ## What changed since yesterday
  <one-paragraph diff summary of any project I touched>

  ## Open questions for me
  <a short list of things you would ask me at the start of the day>

Do not modify anything outside daily-notes/. Do not read inbox/ files.
Do not append to MEMORY.md from this scheduled task.

The schedule runs only when Claude Desktop is open. If your laptop sleeps overnight, the morning task fires when the machine wakes — usually within a minute of opening the lid. There is no background daemon. Plan for that.

What to schedule, what not to

Good fit: repeating summaries of folder contents — morning briefing, end-of-week recap, monthly archive review.
Good fit: low-stakes drafts that benefit from a head start — a weekly update template, pre-filled meeting notes.
Bad fit: tasks that touch inbox/. That folder is a safety quarantine — always reviewed by you first. See the Safety page.
Bad fit: anything that writes to MEMORY.md on a schedule. Memory should be earned during interactive sessions, not bulk-appended by an unattended job.

Next — Read before connecting → Going further →