AI chat apps keep what you tell them in three different places, and mixing them up leads to odd answers. There is the conversation you are in now, a short list of facts the app remembers across chats, and a separate workspace for one job. Choose the right one on purpose, so last weekend’s hobby chat does not end up drafting Tuesday’s board deck.
Say you open ChatGPT on Monday morning, still thinking about the youth soccer game you watched on Saturday. Your Tuesday operations review needs 14 slides, so you type “draft the board deck from the brief I pasted yesterday.” The reply opens with “Congrats on the Riverside U10 Lightning win!” and then builds slides around a brief the chat has never seen. You started a fresh chat because the sidebar looked messy, and the brief you pasted on Sunday night is still sitting in the old thread.
Three places, not one pile
People talk about “ChatGPT remembering me” as if the product had one brain. It does not. You are looking at three, and sometimes four, places that keep different things for different lengths of time. Button labels will change, but the jobs these places do stay stable enough to teach.
History is the conversation you are in, plus the sidebar list of past conversations. Memory is a smaller set of facts the product may keep and reuse in later chats. A Project (or a Gem, or a similarly named space) is a workspace you build for a job, made of files plus standing instructions. Temporary is the “please do not file this” option. Some products also offer a “use my past chats” switch that is separate from the Memory list, so treat it as Memory-adjacent until you read the vendor help page the week you care.
Rule of thumb: If you would not put the sentence on slide 1 of the deck, do not leave it in Memory. If you need the brief next Tuesday, put it in a Project, not in a thread you will lose in the sidebar.
Chat history is this thread
History is the scroll. You asked, the model answered, you pushed back, and that stack is the thread. Most products also keep a searchable sidebar list of threads, and some let you archive them. Closing the tab usually does not delete a thread when you are signed in, and the same account on your laptop often shows the same list. That is convenient, but it is not a filing cabinet.
Your brief lived in Sunday night’s thread, six pages you pasted in after a rain delay. On Monday you started a new chat, and the brief did not travel with you. The soccer team did, because that was stored in Memory. You then spent eight minutes hunting for the old thread under its auto-generated title, “Carpool and slides.” The brief was there, but the new chat had never seen it. A new chat starts empty of anything you pasted before unless Memory, a Project, a Gem, or a “reference past chats” feature feeds it in. Long threads also get fuzzy, so either stay in the thread or move the source into a Project. When the product lets you, name the thread yourself, because “Q3 ops deck, 2026-09-08” beats “Carpool and slides.”
Memory is facts that travel
Memory is the set of sticky notes the product writes about you and then brings into later chats. OpenAI’s Memory FAQ is the clearest vendor page on this. It says that when Memory is on, ChatGPT may keep useful context from chats, files, and connected apps so you repeat yourself less. You can turn it off, and you can open a Memory summary and delete items. The same FAQ is honest that the summary is not a full dump of everything the system might use, so do not treat the list as a complete audit log.
Other products work in a similar way. Gemini has shipped personalization that can draw on past chats (look for Personal Context, though the names change), and Gems are a different feature. Claude’s obvious job workspace is Projects, and whether your Claude plan also shows a global Memory-style list is a Settings check to make this week. Grok keeps renaming similar controls. If you cannot find a list of saved facts, assume the product may still be personalizing.
Memory should hold things that are true in every chat: your role, your voice, your tools, and your standing constraints. Good examples are “I write ops decks for a vice president (VP) who wants unit cost on slide 3” and “Short sentences, no sports talk.” Memory should not hold your kids’ names and teams, your home address, health details, a client’s nickname, or passwords. Suppose your list has 11 saved items and you open it for the first time. Ten are useful notes and one is a leak, and the leak takes 40 seconds to delete, although it has already reached your draft. Some ChatGPT accounts also have a “reference chat history” switch (the wording changes) that can pull from old threads even when a fact is not in the Memory summary. If Sunday should never flavor Monday, archive that thread, turn the switch off, or use Temporary for the Sunday stuff.
Projects are a workspace for one job
A Project is a container you build for a job. You give it a name, standing instructions, and files, and new chats inside it should see those files and instructions. Memory tries to follow you everywhere, while a Project stays in its box, or at least it should.
Claude Projects are workspaces with their own chats, uploaded files, and instructions, and Anthropic’s help page “What are projects?” explains them. Free accounts can create only a few. As of October 2, 2026, Anthropic’s help center and pricing page list a limit of five for free users. ChatGPT Projects have the same shape, and you can pick default memory, where global notes can bleed in, or project-only memory. The default is how a soccer team walks into a client folder, so use project-only for client work when the product offers it. Temporary chats cannot be dropped into a project. Gemini Gems wrap instructions plus knowledge files, and a new Gem chat often reloads those original files, not last Tuesday’s conversation. Keep to one job per workspace. A Gem named “Work” is how a client nickname and a soccer team end up side by side. Team plans that share a project are a way for private material to leak, so if the workspace has a client brief in it, check who can view it.
Your fix for the deck is not “turn Memory off forever.” Instead, create a project named “Tuesday ops review.” Upload last year’s 12-slide PDF and this year’s 6-page brief, once you find the Sunday thread. Then paste in instructions: the audience is the VP of Operations, the talk runs 12 minutes, unit cost goes on slide 3, no family details, and if a number is missing, say “not in the brief.” Now the soccer team has nowhere to stand.
Temporary chats exist, and the names change
Several products now offer a chat that tries not to join your history or Memory. ChatGPT calls this Temporary Chat. OpenAI’s Temporary Chat FAQ says those chats do not appear in your history, do not use or create memories, and are not used to train models. For safety, OpenAI may still keep a copy for up to 30 days, and custom instructions at the account level may still apply. The button has lived as a pill in the top right of a new chat, but check, because interfaces move. Gemini has shipped a Temporary Chat with a similar pitch, and public write-ups have described shorter hold times than 30 days, so open Google’s help page the week you need it. Claude and Grok have offered incognito-style modes at various times. If you cannot find the control, do not assume it exists.
Temporary is the right choice for Sunday carpool, a medical “what does this word mean,” or a messy paste. It is the wrong choice for the board deck you need the next morning. Your Sunday chat should have been Temporary and the deck should have been a Project, but you used one ordinary chat for both because it was late and raining. That is how one of 11 memories went wrong. Temporary is not a legal eraser, since a 30-day vendor hold is still a hold. If a file cannot go to a US vendor at all, running a local model is the topic of a different series.
Which one for which job
Use the table below as a map, then confirm each row in the product. The answers assume you are signed into the same account. A work laptop with a different login counts as a different person to the vendor, and that identity split belongs to the setup series.
| Drawer | Survives closing the tab? | Shared across devices? | Good for | Danger |
|---|---|---|---|---|
| History | Usually yes, as that thread in the sidebar | If you are signed in, often yes | Continuing this conversation | You think the brief is filed. A new chat is empty. |
| Memory | Yes, until you delete the item or turn Memory off | Yes. That is the point. | Stable job, voice, tools | Sunday soccer in a Monday deck. Client names that travel. |
| Project / Gem | Yes. The bucket, files, and instructions stay. | Signed in, on that product, yes | A recurring job with source files | Default Memory bleed. One bucket for every client. |
| Temporary | Usually no lasting sidebar item. Vendor may hold a short copy. | Often no thread to reopen later | One-off, sensitive, or messy paste | You needed the transcript the next morning and it is gone. |
Here is the same Monday with a Memory audit. This is your list of 11 saved items, with the one that leaked called out, using toy labels.
| Saved item (as you read it) | Keep? | Move instead |
|---|---|---|
| Writes ops decks for the Tuesday review | Keep in Memory | |
| Short sentences, no filler, no sports talk | Keep in Memory | |
| VP wants unit cost on slide 3 | Keep, or put in project instructions | Project instructions if it is only this deck |
| Kid plays for Riverside U10 Lightning, Saturday 8am, teal | Delete | Temporary, if she chats soccer at all |
| Client “Harbor” nickname from a Slack paste | Delete | The Harbor project only, never global Memory |
The other six of the 11 items were tool names, a preferred table layout, and “do not invent numbers,” all of which are fine. The Harbor nickname was item 11, and you spotted it on a second pass. Soccer was the loud miss, but Harbor would have been the expensive one in a shared slide. Paste the block below into a note and then into a project’s instruction box, changing the brackets to match your job.
What lives where (one vendor)
Vendor: [ChatGPT / Claude / Gemini / Grok]
Account email: [work address]
Date checked: [YYYY-MM-DD]
HISTORY
- This thread only. Yesterday's brief is in the old thread, not a new chat.
- Sidebar title I will use: [Q3 ops deck, 2026-09-08]
MEMORY
- Toggle: [on / off]
- Screen I opened: [Settings, Memory or Personalization]
- Count of saved items: [11]
- Deleted this week: [Riverside U10 Lightning] [Harbor nickname]
- Never put here: kids, health, home, client names, passwords
PROJECT / GEM / SPACE
- Name: [Tuesday ops review]
- Memory setting if offered: [project-only, not default]
- Files in the bucket: [brief.pdf] [last-year-12-slides.pdf]
- Who can view this (team plans): [only me]
TEMPORARY
- Use for: Sunday carpool, medical words, messy one-off paste
- Do not use for: the deck I need at 7:40am
PROJECT INSTRUCTIONS
Job: Board deck for the Tuesday ops review.
Audience: VP of Ops, 12 minutes.
Voice: short sentences. No sports talk. No family details.
Must include: unit cost on slide 3, if the brief has it.
Never mention: kids, teams, home, personal calendar, client nicknames from other work.
Cite only files in this project.
If a fact is missing, write "not in the brief" instead of guessing.This block forces a count, and eleven is a number you can finish in one sitting. “I think Memory is off” is how the Lightning jersey gets into slide 1. If your vendor has no Project feature yet, keep the instruction block in a document and paste it at the top of each work chat. That is a little ugly, but it is clear, and it beats hoping history filed things correctly.
Common mistakes
- Starting a new chat every morning and wondering why the brief “forgot you.” That was history. Open the old thread or put the brief in a project.
- Leaving Memory on and never opening the list. Eleven items were saved, nobody had looked once, and the one bad item showed up in a draft.
- Putting client work in the default Memory pool, then opening a Project that inherits it. Prefer project-only memory when the product offers it.
- One Gem or Project named “Work” that holds every PDF you own. Names will mix, so keep one job per workspace.
- Drafting the real deck in Temporary, then losing the transcript. Temporary is for the rain delay, not for the VP review.
- Treating Temporary as “the vendor never saw this.” Read the FAQ, because a 30-day safety copy is still storage. Attachments in a regular thread are also not a Project knowledge base, and a later post in this series covers where the file went.
How to practice this week
- Open your main chat product and find Memory (or Personalization, or the closest list of saved facts). Count the items and write the number down. Delete anything a VP should not see.
- Find Temporary (or the incognito-style control). Send one throwaway question, and confirm it does not land in the sidebar the way a normal thread does.
- Create one Project or Gem for a real job you will do twice. Put two files in it and paste the instruction block above. If the product offers project-only memory, turn that on.
- Rename one old sidebar thread so you can find last week’s brief without scrolling for “Carpool and slides.”
The next post is Uploading docs safely, which covers what to strip out before a file goes into the workspace. Learn lists the other paths, and the chooser is still the place to decide which product to use.
Quick recap
- History is this thread and the sidebar list. It is not a filing cabinet.
- Memory is facts that travel. View them, count them, and delete the ones that do not belong. In the story, 11 items and 1 that should never have been there.
- A Project or Gem is a workspace for one job, with files and instructions. It is not Memory.
- Temporary is the “do not file this” option. Names change, and vendors may still hold a short copy.
- Put Sunday soccer in Temporary and the brief in a project, and keep Memory for your job and your voice.
Series notes
This is Part 3 of Prompting for everyone (history, memory, projects). Related: Files and uploads, Iterate when wrong.
Sources
- OpenAI: Memory FAQ (how Memory works, view and delete, Temporary as the opt-out)
- OpenAI: Projects in ChatGPT (project files, chats, memory settings; check the current help page)
- OpenAI: Temporary Chat FAQ (no history item, no Memory, possible 30-day safety copy)
- Anthropic: What are projects? (workspace, knowledge files, instructions, free-user cap; check the current help page)
- Google: Tips for creating custom Gems (instructions plus knowledge files)
- Google Gemini app (confirm Temporary Chat and personalization labels in the live UI)
- Grok (confirm any Memory or space labels the week you use it)
- AMS: AI for files and office work
- AMS: AI setup from zero
Keep going
Same lessons in your feed
Short diagrams, hooks, and weekly tutorials on Substack, Instagram, X, and Facebook.
