The Cowork Academy by Jeff Su
Free interactive guide

How Does AI Memory Actually Work?

Memory is on, and your AI still asks what you're working on. The 3 levels below explain exactly why, and how far you can take it.

Level 1 · Global Memory

Watch your AI learn you today, then forget you overnight.

Global memory is the account-level notes your AI writes about you once the memory feature is on, and they apply across every chat.

Click “Start the session” and watch the right-hand panel
Yesterday · 4:52 PM
ChatGPT Gemini Claude

Works the same in all three.

The AI learns:
  • Nothing yet. Start the session.

Yesterday, 4:52 PM. A normal work session.

This is deliberate.Anything saved here gets dragged into every single future chat:

Global memory · one file for your whole life
Works in operations
Likes short, punchy writing
“Pitch moved to Oct 6” (you forced it in)
Work deck
Kids' school
Side business
Golf trip

Everything here flows into every chat you open.

Think about how many hats you wear:

Information that's useful in one context is noise in another, so global memory errs on the side of remembering less. It keeps the high-level you, because it has to work across your entire life.

You can force it mid-chat: say "Update your memory: the pitch moved to October 6" and it works. But you have to remember to ask every time, and that entry follows you into chats that have nothing to do with the pitch. Connectors can pull in documents, but they don't know what happened inside yesterday's session.

Check your own right now: in ChatGPT it's Settings › Personalization › Memory › Manage. Claude and Gemini have their own versions, and they behave the same way. You'll find your role and writing preferences, and almost nothing about any specific project.
ChatGPT
ChatGPT memory settings with Enable memory on and the Manage button

Settings › Personalization › Memory

Claude
Claude memory settings showing generated memory: preferences and profile

Settings › Memory

Gemini
Gemini Personal Intelligence settings with Memory enabled

Settings › Personal Intelligence

Level 1, explained in plain English

Global memory is your AI writing account-level notes about you that apply to every chat. That's useful for the high-level details that should follow you across all the roles in your life, but it will never hold enough specific context for real work, like an ongoing project.

Level 2 · Project Memory

A Project remembers your work.The AI still holds the pen.

The Projects feature draws a boundary around one workstream, so the AI can keep specific memory about the actual work.

This is the minimum setup for real work, and it has one catch.

Click through the project session and watch the Memory card
Projects / Micro-iPhone Leadership Pitch
Micro-iPhone Leadership Pitch

Prepare a high-stakes Micro-iPhone pitch for the Nutella-Cookie leadership team.

Decks drafted, decisions made, and a calendar invite pasted with 10 names. ChatGPT Projects and Gemini Gems work the same way.

Memory🔒 Only you

Written by the AI, updated as you chat.

  • Start the session to see what it wrote.

A real Project, one week in.

Everything so far has the same root issue: the AI is still in control. It decides what gets remembered, where it lives, and when it gets written.

The fix at this level is manual: “Update project memory: John Turnip is confirmed along with Tim Cookie.” It works, but you have to catch the gap first, and you usually find out when an answer comes back wrong.

Claude's Manage project memory modal, which explains that Claude regenerates project memory every evening from your past chats
The real thing. Claude regenerates this Project's memory from your chats, every evening. Notice who decides what makes the cut.

Level 2, explained in plain English

Projects give the AI a tighter boundary, so it writes more specific memory about one workstream, and every chat inside still inherits your account-level memory from Level 1. The catch: the AI still picks what to keep, so answers can be confidently outdated.

Level 3 · Your Own Memory System

Memory stops being a black box and becomes files you control.

Your AI reads plain-text files at the start of every session and updates them as you work. You decide what gets saved. The AI does the upkeep.

Fill in three fields, and your first memory file writes itself

The file updates as you type. Nothing you enter is stored or sent anywhere; this page has no server.

MEMORY.md

      
If you want the full Level 3 system

I already built the system

The MEMORY.md file you just created is one small part of the system I use every day.

The Cowork Academy gives you the working template, 10+ pre-built workstations, and the step-by-step course to make the system yours.

Claude Cowork
MEMORY.md Meetings HQ Comms HQ Personal Finances
Your week, ready:
Mon · Micro-iPhone polish. Brief drafted.
Tue · Leadership sync. Talking points ready.
Fri · Card payment due. Reminder set.
2 replies drafted, waiting for your OK.

One sentence in. Your whole week prepped, because the system already knows you.

Two weeks from now, Monday morning starts with a draft already waiting.

Here's what The Cowork Academy changes in practice. Same request on both sides. On the right, the context loads when it's needed, from files you control.

Drag the handle to compare
Claude Cowork
Context
Reply to Claude…
Before
Task

Go through my inbox and draft replies to anything that needs one.

Context

I run a small business. Vendor mail belongs to Business Operations and client work to Product Marketing. I am mid-launch, so anything about it is high priority. My co-founder handles invoices, so pass those along to her.

Constraints

Never send anything without showing me first. Keep every reply short. Do not promise a date I have not confirmed.

Preferences

No em dashes. Bullets over paragraphs. Sign off as Best, Jeff. Warm but direct, never stiff or corporate. If a thread has gone quiet for a week, nudge it. Anything from a client outranks anything internal.

Retyped every single session

I’ve reviewed your inbox. Here is a summary of the items I found, and some messages that may require your attention. Let me know how you would like to proceed.

Generic. Not your voice. Needs corrections.

Nothing loaded
After

Go through my inbox.

Pulls your contextComms HQVoice Principles.mdMEMORY.md
Done. 0142 newsletters labeled and archived, invoices passed to your co-founder.3 need you this week:• Sarah (Acme): the timeline slipped. 02Reply drafted, short, no dates promised.03Legal: contract redline, Thursday.• Byron: intro request, can wait.
Relevant
Sounds like you
Ready to send
01Gmail triage rules
Comms HQ
02Communication preferences
Voice Principles.md
03Active projects
MEMORY.md

Same request, same AI. The difference is three files the system loaded on its own.

Level 3, explained in plain English

You write the rules for what gets saved and where, and the AI does the upkeep. Total control, minimal work. The honest downside: a real system takes effort to set up, and it works differently from the chat window you're used to.

Recap

Which level do you need?

Level 1

Global Memory

Automatic and account-wide. Keeps the high-level you, loses the actual work. Fine if AI is an occasional assistant.

Level 2

Project Memory

Automatic with a tighter boundary. The floor for real work, but the AI still picks what to keep, so answers can be confidently outdated.

Level 3

Your Own System

Plain-text files you control, updated by the AI as you work. Total control, minimal upkeep, real setup effort.

Prefer to start free?

Build the foundation in 7 days

The MEMORY.md file you created above is step one. The free Cowork Toolkit helps you build the foundation yourself over seven days, one clear layer at a time.

Day 1

Install the two root files

Give Cowork a rulebook and a place to keep current context.

CLAUDE.mdMEMORY.md
Day 2

Teach Cowork your voice

Extract the patterns that make drafts sound like you.

Voice Principles.md
Day 3

Build Comms HQ

Create your first workstation for email, communication rules, and follow-ups.

Comms HQ/
Day 4

Add Personal Finances

Build a second workstation, then teach Cowork where different work belongs.

Personal Finances/routing
Day 5

Create your first project

Give one focused piece of work its own context inside a workstation.

Subscription Audit/
Day 6

Turn on the audit loop

Use memory, routing, and session audits to keep the system improving.

memory auditrouting auditsession audit
Day 7

See the full system

Review what you built, then decide whether you want the faster path to all 15 workstations.

By Day 7, you'll have two workstations, one project, and a starter audit loop working together.

The Cowork Academyby Jeff Su
© 2026 The Cowork Academy