Kestrel Exams kestrel logo KESTREL EXAMSPRACTICE TO PASS
Using AIFree starter prompt17 min readChecked October 2026

Give Your AI a Memory

Most AI chats start from scratch. Download one starter prompt on your computer, give it to Claude or ChatGPT, and it sets up a simple memory system with you: six short files your AI reads at the start of every chat and updates at the end.

Download the starter prompt ↓Back to the course

Start here: why your AI forgets#

An AI assistant does not remember your work the way a coworker does. Much of what you told it last week is gone when you open a new chat, unless something carries it forward. There are two ways to carry it forward, and the best setup uses both.

The app's built-in memoryYour own memory files
What it isThe app saves some facts about you and brings them into later chats.A small folder of plain text files that you and the AI keep up to date.
Who decides what is keptMostly the app, though you can view, edit and delete it.You. You can read every line.
Works in other appsNo. It stays in that app.Yes. Any AI that can read a file can use them.
Best forHow you like answers written.Projects: what is done, what is next, what was decided, what went wrong.

Kestrel Exams is a small study site built by one person working with an AI assistant, and this is the system we use ourselves, simplified. This page explains it. The starter prompt below does the setup with you: you upload it to your AI app, and the AI walks you through each step.

How the starter prompt works#

  1. On a computer, download the starter prompt. It is one text file. You need a Windows PC or a Mac for setup: you will save files into a folder and change app settings, which a phone cannot do well.
  2. Open a new chat in Claude or ChatGPT, on the desktop app or the website. Attach the file with the paperclip or + button (or open the file and paste its text in), then type Follow the instructions in this file.
  3. Answer its questions. It asks which app, computer and plan you have, then helps you make a project folder, interviews you to write your six memory files, and walks you through your app's settings. It finishes by testing that everything works. Allow about 20 minutes.

On a phone? Open this page on a computer instead. Setup means saving files into a folder and changing app settings, which needs a computer.

AI Memory Starter Prompt

Use a computer (Windows or Mac). One text file (Markdown, about 18 KB). Upload it to a new chat in Claude or ChatGPT and type Follow the instructions in this file.

Download the starter prompt

The file is plain text: you can open it in any text editor and read every instruction it gives the AI before you use it. View it in your browser.

The six memory files#

Each file has one job. Keeping the jobs separate is what stops the files turning into one long, confusing note.

FileIts jobHow it changes
START_HERE.mdInstructions for the AI: what to read first, how to work with you, how to save progress.Rarely.
PROJECT_STATE.mdThe current truth in one page: where things stand and what is next.Old lines are replaced, so it stays one page.
TASKS.mdThe to-do list, with who does each task.Boxes get ticked; finished tasks are archived monthly.
CHANGELOG.mdA dated diary of each session and what you decided, in your words.New entries go on top. Nothing is rewritten.
MISTAKES.mdWhat went wrong and the rule each mistake taught.New entries on top. The AI reads it before working.
PREFERENCES.mdHow you like to work: length, tone, when to ask first.Whenever you want something done differently.

The .md ending means Markdown: plain text with a few symbols for headings (#) and lists (-). Any text editor can open them: Notepad on Windows or TextEdit on a Mac.

Five rules that keep the files trustworthy

  1. Newest on top, and every entry dated as YYYY-MM-DD, such as 2026-10-21.
  2. Decisions in your own words. Your exact words beat the AI's summary of what it thinks you meant.
  3. One source of truth. Keep one copy of the folder that counts. If the app's memory and the files disagree, the files win.
  4. Add lines; never retype whole files. When an AI rewrites a long file from memory, lines go missing without anyone noticing.
  5. Keep a backup in a second place: a cloud drive, a USB stick or an emailed copy. Back up before letting any AI edit the files.
From our own work: the deleted records

While updating four important record files, our AI assistant wrote a small helper in Windows PowerShell, a command tool built into Windows, and named it Rd, meaning "read". Rd was already PowerShell's built-in shortcut for delete. The four files were permanently deleted, not sent to the Recycle Bin. An older backup and the parts the assistant had read minutes earlier rebuilt most of them; about a week of entries had to be rewritten from notes and marked as reconstructed. The rule we kept: keep backups, and back up before rewriting anything important.

Setting up the Claude desktop app#

Checked against Anthropic's help center in October 2026. Menus move; if one is not where described, search the help center for the feature name.

1. Install it

Download the app from claude.ai/download. It runs on macOS 13 (Ventura) or newer and Windows 10 or newer. Open it and sign in. [1]

2. Tell it how you like to work (all chats)

Click your initials in the lower left corner, choose Settings, and find Instructions for Claude. Paste a few lines from your PREFERENCES.md: length, tone, how much to explain. Keep it short. Project details go in the project, not here. [2]

3. Check its built-in memory

Claude's Memory is on by default on the Free, Pro and Max plans. On Team and Enterprise plans, an owner decides, and memory stays off for each member until they turn it on. The controls are under Settings > Memory: you can see what it remembers, edit or delete an item, pause memory, or reset it. Each project also has its own memory. Inside a project, Claude uses that project's memory plus your general memory, but not memory from other projects. To keep a project fully separate, turn off Use account memory under Memory on the project's page. [3]

4. Make a project for your work

A project is a workspace in the app that keeps related chats together with shared files and instructions.

  1. Open Projects in the left sidebar and click + New Project. Name it after your folder. Claude cannot see the name or description, so do not put instructions there. [4]
  2. Click Set project instructions, paste the standing instructions from the starter prompt, and click Save instructions. These apply to every chat in the project. [4] If your plan does not show this option, skip it: START_HERE.md carries the same instructions, and the start-a-session prompt tells Claude to read it.
  3. In the project's knowledge area, click + and add your six memory files. Claude reads project files in every chat in the project; apart from the project's memory, chats in a project do not share what was said. [4]
  4. Start each chat inside the project with the start a session prompt.
One catch: uploaded files are copies

When the files change on your computer, the copies in the project do not. After each "save our progress", replace the changed files in the project: remove the old copy and add the new one. Free accounts can have up to five projects. [4]

5. On a paid plan: let Claude update the files itself

On the Pro, Max, Team and Enterprise plans, the desktop app can read and write files in a folder you connect, so Claude can update your memory files itself at the end of a session. It can only use folders you have connected, and only while the desktop app is open. Anthropic's help article on connecting a folder shows the current steps. [5][6] If you work this way, stop uploading copies to a chat project, so there is only one version of each file.

  • Connect your project folder only. Anthropic suggests a dedicated working folder rather than broad access, and caution about sharing sensitive files such as financial documents, passwords or personal records. [6]
  • Claude asks for your permission before permanently deleting any file. Read those requests. [6]
  • Even so, keep the backup from the five rules above.

6. Privacy settings worth a look

  • Model training: under Settings > Privacy, the Help improve our AI models setting decides whether your chats may be used to train models (Free, Pro and Max). [7]
  • Incognito chats: click the ghost icon on a new chat. These are not saved to your history or memory and do not use your memory. They only work outside projects. [8]

Setting up the ChatGPT desktop app#

Checked against OpenAI's help center in October 2026. OpenAI's new desktop app combines Chat, Work and Codex; the earlier app is now called ChatGPT Classic and can still be used. Menus may differ slightly between them. [9]

1. Install it

On a Mac, download it from chatgpt.com/download. It needs macOS 14 or newer on Apple Silicon (M1 or later) or an Intel processor. On Windows, get it from the Microsoft Store; it needs Windows 10 or newer. Some OpenAI pages say the desktop app is for paid plans, so check what your plan includes. If your plan does not include the app, the same settings and projects are at chatgpt.com in a web browser. [10][11]

2. Tell it how you like to work (all chats)

Open Settings > Personalization, turn on Enable customization, and paste a few lines from your PREFERENCES.md into Custom Instructions. The limit is 1,500 characters on the Free and Go plans and 5,000 on Plus, Pro and business plans. [12]

3. Check its built-in memory

Go to Settings > Personalization > Memory. Reference saved memories lets it use facts it has saved about you; Reference chat history lets it draw on past chats. You can ask ChatGPT to forget something, or delete it in the same settings. Deleting a chat does not necessarily delete a memory saved from it. What is available depends on your plan and region. [13]

4. Make a project for your work

A project keeps related chats together with shared files and instructions.

  1. Select New project in the sidebar and name it after your folder. When asked, choose project-only memory so chats outside the project do not mix in. You can change this later under Project settings > Memory; changes can take a few hours to apply. [14]
  2. Open the project's ••• menu, choose Project settings, and paste the standing instructions from the starter prompt as the project instructions. They apply only in this project and override your Custom Instructions there, so inside the project your PREFERENCES.md file does that job. [14]
  3. Add your memory files to the project. The Free plan allows 5 files per project, so on Free leave out START_HERE.md: the standing instructions already cover what it says. Go and Plus allow 25, higher plans 40. [14]
  4. Start each chat inside the project with the start a session prompt.
The same catch: uploaded files are copies

After each "save our progress", replace the files that changed in the project, so the next chat reads the current version.

5. On a paid plan: working with a folder

In the new app, Work (a mode for longer, multi-step tasks) can open a local folder on eligible paid plans, so ChatGPT can work with your files directly. OpenAI advises granting access only to the files a task needs, and notes that messages and task context may be stored in the cloud even when the work runs on your computer. [15]

6. Privacy settings worth a look

  • Model training: under Settings > Data controls, the Improve the model for everyone setting decides whether your chats may be used for training. Turning it off does not delete saved chats. [16]
  • Temporary chat: choose Temporary in a new chat before your first message (a Personalized temporary chat can still use your existing memories). These do not create or update memories, stay out of your history and are not used for training, though they may be kept up to 30 days for safety. [17]

Add-ons that help: skills, plugins and connectors#

Both apps let you add extra abilities. A few of them fix the biggest weakness of this system: files uploaded to a project are copies that go stale. Here is what helps, with the plan and risk of each.

Add-onWhat it does for your memory filesPlanRisk
Memory keeper skill (Claude, our free download)Teaches Claude the start, save, mistake and tidy routines, so you can just say "save our progress".All Claude plansLow: you can read every line
Google Drive connector (Claude)Keep the files as Google Docs; a Claude project always reads the latest version.All Claude plansLow
Filesystem extension (Claude desktop)Claude reads and edits the real files in your folder: no copies at all.Desktop appMedium: it can change files
Plugins (Claude)Bundles of skills and connectors for a kind of work. Optional.Paid Claude plansDepends on the plugin
Apps (ChatGPT)Let ChatGPT search services such as Google Drive. Drive files added to a project do not stay in sync.Varies by plan and regionLow to medium

Claude: the memory keeper skill

A skill is a folder of instructions Claude loads when a task calls for it. Skills work on every Claude plan, including Free, but they need code execution turned on. [18]

  1. Turn it on: Settings > Capabilities, then switch on Code execution and file creation. On Enterprise plans an Owner must turn skills on first; on Team plans they are on by default. [18]
  2. Download our skill: memory-keeper-skill.zip. Do not unzip it. If your Mac unzipped it for you (Safari does this), right-click the memory-keeper folder and choose Compress to make the zip again. It holds one folder with one instruction file, SKILL.md, which you can open and read first. [19]
  3. Upload it: Customize > Skills, click +, then + Create skill and Upload a skill, and choose the zip. Switch it on in the same list. [18]
  4. Test it: in your project, type start a session. Claude should brief you from the files, and the steps it shows should mention memory-keeper.

Claude also ships ready-made skills for Excel, Word, PowerPoint and PDF files, and a skill-creator skill that helps you write your own. [18][19] Anthropic's advice: only install skills from sources you trust, and read a skill from someone else before switching it on. [18]

Claude: keep the files in Google Drive so they never go stale

The Google Drive connector is available on every Claude plan. Connect it under Customize > Connectors: choose Discover, search for Google Drive, click Connect to Claude, and sign in to approve. On Team and Enterprise plans an Owner may need to add it first. [21][22]

Then make each memory file a Google Doc with the same name, such as PROJECT_STATE.md (paste in the text) and add the Docs to your project with Files > + > Drive. Anthropic says Google Docs added to chats and projects sync from Drive, "so you're always working with the latest version." This does not work in shared projects. [22] You still paste each "save our progress" update into the Docs yourself, but you never re-upload anything.

Claude: the Filesystem extension (for comfortable users)

A desktop extension is a small add-on that runs on your own computer, with the same permissions as your other programs. [20][21] Anthropic's Filesystem extension lets Claude read and write files directly. [24]

  1. In the desktop app, go to Settings > Extensions and choose Browse extensions. Find Filesystem (marked as made by Anthropic) and click Install. [23][24]
  2. When it asks which folders Claude may use, choose only your project folder, never your whole drive or Documents folder. If you cannot see where to choose folders, stop and do not use this extension.
  3. Back up the folder first. Claude can now change real files, so read what it plans to do before you approve it.
  4. Stop uploading copies of the files to a Claude project, so there is only one version of each file.

Claude: plugins

A plugin bundles skills, connectors and helpers into one package. Plugins need a paid plan (Pro, Max, Team or Enterprise). Browse them under Customize > Plugins > Discover and click Add. Anthropic's own plugins cover areas such as marketing, finance, legal, design and data, and there is a plugin for building your own. Only install plugins from sources you trust: some run programs on your computer with your permissions. [20]

ChatGPT: apps and skills

  • Apps connect ChatGPT to services such as Google Drive, GitHub or Notion. Open Plugins or Apps, whichever your account shows, and choose the app. Click Install plugin if you see it, then Connect if asked, and sign in. What is available depends on your plan and region. Data you share with an app is handled under that app's own terms. [25]
  • Google Drive files added to a ChatGPT project do not stay in sync, so keep replacing the copies after each save. ChatGPT can still search your Drive through the app. [14]
  • Skills in ChatGPT are for Business, Enterprise, Healthcare and Edu plans, under Plugins > Plugin Directory > Skills. [26]
  • Avoid building your memory system on a custom GPT: OpenAI plans to retire custom GPTs (scheduled for December 11, 2026, though your account may get a different date) and to help creators move them to plugins. [27]
Before you add anything

Install only from sources you trust, read the permissions you are asked to approve, and choose "always allow" only for tools you trust completely. Text inside a connected document or website can try to give the AI instructions; if it starts doing something you did not ask for, stop it.

The two habits that make it work#

The files only help if they stay current. Start every work session with the first prompt and end it with the second. Both are in the starter prompt file too.

Start a session

Read START_HERE.md first (if it is there), then the other memory files. Then brief me: where we are, the top three next steps with your recommendation, and anything you need from me. To show you read them, quote the heading of the newest entry in CHANGELOG.md and the first unticked task in TASKS.md.

Those two lines change every session, so the AI can only quote them if it really read the files.

Save our progress

Save our progress.
- CHANGELOG.md: a new dated entry at the top. What we did, decisions I made (in my words), files changed.
- TASKS.md: tick what is done, add what is new.
- PROJECT_STATE.md: replace anything that is no longer true. Keep it under one page.
- MISTAKES.md: anything that went wrong today, with the rule it teaches.
If you can edit the files directly, show me each change before you save it. If you cannot, give me only the new or changed lines for each file and tell me exactly where they go. Do not retype whole files.

If you uploaded the files to a project, replace the changed ones there afterward: the project holds copies. The file also includes prompts to log a mistake, tidy the files monthly, hand off to a new chat, and check what the AI remembers.

From our own work: the mistakes log

We asked our AI assistant to keep a running log of its own mistakes and the rule each one produced, newest first, and to actually apply them. Several later entries say "applied successfully." When you work with AI over time, a list of how it fails you is the most useful prompt you own.

Privacy, safety and troubleshooting#

  • Check your data settings (see the app sections above) and decide whether your chats may be used to train models.
  • Keep secrets out of the files: passwords, card and bank numbers, ID numbers, other people's private details. Anything in these files is sent to, and may be stored by, the AI company.
  • Work and school rules first. Check your employer's or school's policy before putting their material into an AI tool, and use the account they approve.
  • Give folder access to the smallest folder that works: your project folder, never your whole drive.
  • Read before you approve. When an app asks to run a command or change a file, read what it will do first.
What happensWhat to do
It answers without reading the files.Ask: "Which files did you read? Quote the newest CHANGELOG.md heading." Check the files are in the project.
It "remembers" something that never happened.Ask: "Show me where in the files that is." If it cannot, correct it and log it in MISTAKES.md.
It quotes old information.The project still holds old copies. Remove them and add the current files.
The files keep getting longer.Use the monthly tidy prompt. Keep PROJECT_STATE.md to one page.
Built-in memory and the files disagree.The files win. Edit or delete the wrong memory in the app's settings.
A very long chat starts losing track.Save progress, then start a fresh chat with the start-a-session prompt.
From our own work: the plan that already existed

Asked to "review and update the plan" for a video series, our AI assistant wrote a new plan from scratch. A detailed 23-module plan had existed for a week, in the same cloud folder the assistant had opened minutes earlier. The rule for the human: point the AI at the documents that matter.

That is what START_HERE.md is for: it points the AI at the documents that matter, every time.

Frequently asked questions#

Does this work with the free versions of Claude and ChatGPT?

Yes. The memory files and prompts work on free plans. Some extras, such as letting the app edit files in a folder on your computer, need a paid plan, and the ChatGPT free plan allows 5 files per project, so the starter prompt has you leave one file out.

Is the starter prompt safe to upload?

It is a plain text file of instructions and blank templates, with no code. You can open it in any text editor and read every line before you use it.

Do I still need the app's built-in memory?

It helps with general preferences, like how long you want answers to be. Keep project details in your own files, where you can read and correct every line.

Will this work with other AI assistants?

Yes. Any assistant that can read an uploaded file can use the memory files. The starter prompt tells the AI what to look for in apps other than Claude and ChatGPT.

Free, with no ads and no account. If this helped you, you can help keep it that way.

Support Kestrel Exams →

Sources#

Every app detail on this page was checked on these official help pages in October 2026. Apps change often; if a menu is not where described, search the app's help center for the feature name.

  1. Anthropic. Install Claude Desktop. Claude Help Center, checked 2026-10-21.
  2. Anthropic. Understanding Claude's personalization features. Claude Help Center, updated 2026-09-16, checked 2026-10-21.
  3. Anthropic. Use Claude's chat search and memory to build on previous context. Claude Help Center, checked 2026-10-21.
  4. Anthropic. How can I create and manage projects? Claude Help Center, checked 2026-10-21.
  5. Anthropic. Get started with Claude Cowork. Claude Help Center, checked 2026-10-21.
  6. Anthropic. Use Claude Cowork safely. Claude Help Center, checked 2026-10-21.
  7. Anthropic. How do I change my model improvement privacy settings? Privacy Center, updated 2026-08-03, checked 2026-10-21.
  8. Anthropic. Use incognito chats. Claude Help Center, checked 2026-10-21.
  9. OpenAI. Moving to the new ChatGPT desktop app. OpenAI Help Center, checked 2026-10-21.
  10. OpenAI. Downloading the ChatGPT macOS app. OpenAI Help Center, checked 2026-10-21.
  11. OpenAI. Using the ChatGPT Windows app. OpenAI Help Center, checked 2026-10-21.
  12. OpenAI. ChatGPT Custom Instructions. OpenAI Help Center, checked 2026-10-21.
  13. OpenAI. Memory in ChatGPT. OpenAI Help Center, checked 2026-10-21.
  14. OpenAI. Projects in ChatGPT. OpenAI Help Center, checked 2026-10-21.
  15. OpenAI. ChatGPT Work and Codex. OpenAI Help Center, checked 2026-10-21.
  16. OpenAI. Data controls in ChatGPT. OpenAI Help Center, checked 2026-10-21.
  17. OpenAI. Temporary chat in ChatGPT. OpenAI Help Center, checked 2026-10-21.
  18. Anthropic. Use skills in Claude. Claude Help Center, checked 2026-10-21.
  19. Anthropic. Create custom skills. Claude docs, checked 2026-10-21.
  20. Anthropic. Use plugins in Claude. Claude Help Center, checked 2026-10-21.
  21. Anthropic. Get started with connectors. Claude docs, checked 2026-10-21.
  22. Anthropic. Use Google Workspace connectors. Claude Help Center, checked 2026-10-21.
  23. Anthropic. Getting Started with Local MCP Servers on Claude Desktop. Claude Help Center, checked 2026-10-21.
  24. Anthropic. Filesystem (desktop extension listing). Checked 2026-10-21.
  25. OpenAI. Connected apps in ChatGPT. OpenAI Help Center, checked 2026-10-21.
  26. OpenAI. Skills in ChatGPT. OpenAI Help Center, checked 2026-10-21.
  27. OpenAI. Custom GPT retirement and migration FAQ. OpenAI Help Center, checked 2026-10-21.

Claude is a trademark of Anthropic. ChatGPT is a trademark of OpenAI. This guide is independent and not endorsed by either company.