Personalizing ChatGPT: Custom Instructions, Base Style and Tone, Memories and Temporary Chat

Version note: This post describes ChatGPT's personalization settings as of 15 November 2026; details on base style and tone, custom instructions, memories, Computer History and Temporary Chat were checked the same day in the ChatGPT documentation (learn.chatgpt.com) and the OpenAI Help Center (help.openai.com). This post is based on the official documentation. Setting names and locations may differ between the web, the desktop app and mobile.

Guess First

Emre is a developer on Bookstore, the online bookstore used as the shared example across this site's series. In one chat he told ChatGPT, "from now on always write code samples in C# 14 and .NET 10, and name classes according to Bookstore's conventions". Over the following days he saw the rule followed in some chats and not in others. The same week, to resolve a customer complaint, he had also pasted the customer's e-mail address and order number into a chat; now he wonders whether that information stayed in ChatGPT's "memory".

What is your guess? Why is Emre's rule applied only sometimes? Where did the rule belong? And is deleting the chat enough to remove the customer's information completely? The answers are in the Help Center's memory article and in one sentence of the memories page on learn.chatgpt.com: memory is a recall layer, not the source of rules that must always be followed.

The Layers of Personalization

There are several separate ways to personalize ChatGPT, and each answers a different question. The "Personalize ChatGPT" page on learn.chatgpt.com (https://learn.chatgpt.com/docs/personalize, 15 November 2026) lists them; most of the settings are under Settings > Personalization on the web:

• Base style and tone: how ChatGPT talks. It changes style and tone, not capabilities.
• Custom instructions: preferences you want followed in every chat. In Codex, these personal instructions are kept in the global AGENTS.md file.
• Writing style: Work learning your style from your own writing in the e-mail, messaging or document apps you connect, so drafts sound like you.
• Memories: context carried from earlier chats into later work.
• Computer History: an opt-in feature in the macOS desktop app that turns your activity in allowed apps and websites into memories and a timeline.

There are two more layers, which later posts cover in detail: project instructions (post four) and skills that come with plugins (post five). A rough way to think about it: custom instructions answer "who am I and how do I want answers", project instructions answer "how is this job done", and memory answers "what did we talk about last time".

Base Style and Tone and Custom Instructions

On the web, you choose a personality on the Settings > Personalization > Base style and tone screen. The documentation stresses that it only changes style and tone, not what ChatGPT can do. For a developer who wants short, direct answers, this setting is a cleaner solution than adding "be brief" to every message.

Custom instructions are free text applied to every chat. According to the Help Center article "ChatGPT Custom Instructions" (https://help.openai.com/en/articles/8096356, 15 November 2026), Free and Go accounts can save up to 1,500 characters, and Plus, Pro, Business, Enterprise and Edu accounts up to 5,000; the limit on paid plans went from 1,500 to 5,000 on 15 July 2026. On the web and desktop, the setting is under Settings > Personalization, together with a switch that turns customization on and off.

On the Codex side, the counterpart of the same idea is a file. Codex keeps personal instructions in the global AGENTS.md file (by default ~/.codex/AGENTS.md); AGENTS.md files in a repository are for team rules, and the file closest to the working directory takes precedence. The details of these files belong to the OpenAI Codex series. What matters here: a preference you write in ChatGPT's custom instructions and the project rules Codex reads live in different places, and neither replaces the other.

Writing style in Work is a separate mechanism. The documentation describes the setup in four steps: open the setup page, connect the e-mail, messaging or document apps ChatGPT should use, enable it with "Use writing style", and ask Work for a draft. Connecting an app alone is not enough; enabling is a separate step. To turn it off, you disable "Reference my writing style" under Settings > Personalization > Writing. It means opening the correspondence in the connected apps to ChatGPT as examples; check your company's policy before connecting a work e-mail account.

Memories: What Does ChatGPT Remember?

The memories page on learn.chatgpt.com (https://learn.chatgpt.com/docs/customization/memories) says up front that there are two separate memories: ChatGPT on the web uses ChatGPT memory, while local Codex clients use a separate local memory store with separate controls. Work follows the memory settings of your account and workspace; it does not use the local Codex memory.

Managing ChatGPT memory is covered in the Help Center article "Memory in ChatGPT" (https://help.openai.com/en/articles/8590148). Under Settings > Personalization > Memory you see a few controls that vary by account: using saved memories (Reference saved memories), using past chats (Reference chat history), a memory summary, and a screen for managing memories. Three details here answer Emre's second question:

Saved memories are stored separately from chat history. Deleting the chat where you shared a piece of information does not delete the saved memory created from it. To remove something completely, you have to delete both the memory on the Manage memories screen and the chat where you first shared the information. According to the article, logs of deleted saved memories may also be kept for up to 30 days for safety and debugging.

If you turn off the setting for using past chats, information remembered from past chats is scheduled for deletion within 30 days; the chats themselves stay in your history unless you delete them, and your saved memories are not deleted. Since 12 June 2026 you can also delete and turn off memory in one step with "Delete and turn off memory" in the three-dot menu on the memory summary page.

Codex's local memory follows different rules and is off by default. In the desktop app you turn it on for the selected machine with "Enable Codex memories" under Settings > Personalization; "Allow memories from tool-assisted chats" decides whether chats that used MCP tools or web search can contribute to memory. Inside a chat, the /memories command decides whether that chat can use existing memories and whether it becomes input for future memories. Codex memories are files in Codex's home directory, and the documentation advises against relying on editing them by hand.

Now back to Emre's first question. A sentence like "from now on always do this" said in a chat becomes, at best, a memory, and memory may not come into play the same way in every chat. The memories page is clear: keep team rules that must always apply in an AGENTS.md file or in documentation in the repository; treat memory as a helpful recall layer, not the only source. In ChatGPT, the counterpart is custom instructions, or the project instructions we will see in the next post.

Computer History

Computer History is the most far-reaching personalization feature in the macOS desktop app, and therefore the one that needs the most careful decision. According to the documentation (https://learn.chatgpt.com/docs/customization/computer-history), it is off by default for Pro, Business and Enterprise users; Pro users can turn it on themselves, while on Business and Enterprise an admin first has to grant access to the member. It requires memories to be on and is not available with an API key.

Once on, it builds an event stream from interactions in the apps and websites you allow, such as clicks, typing, keyboard shortcuts and app switches, plus the text that macOS exposes through its accessibility system, and periodically turns this into text summaries and local memory files. It does not take screenshots and does not record the microphone or system audio; private-mode browsing is never included. You can pause collection from the menu bar, exclude apps and websites, delete individual items in the timeline, or clear all history. Deleting Codex memories also clears this history.

The practical rule for the Bookstore team: before turning it on, add the apps where you work with customer data (the support panel, order management, e-mail) to the exclusion list. Off by default is a good start; the decision to turn it on should belong to the team and the admin, not one person.

Temporary Chat

Some chats you want neither in your history nor in memory. According to the Help Center article "Temporary chat in ChatGPT" (https://help.openai.com/en/articles/8914046, 15 November 2026), a Temporary Chat does not appear in history, does not create or update memories, and is not used to improve OpenAI's models while it stays temporary. OpenAI may still keep a copy for up to 30 days for safety purposes.

You choose one of two options before the chat starts. A personalized Temporary Chat can use your existing memories, custom instructions and plugins but does not create new memories; the unpersonalized option uses none of them. The choice cannot be changed after the chat starts. If you save a Temporary Chat, it becomes a regular chat and from then on follows your account's personalization and model improvement settings. Model training settings are covered in detail in the tenth post; what matters here is that Temporary Chat is not a "deletion guarantee" but an option not to write to history and memory.

Bookstore: Which Rule Goes Where?

Let us put Emre's rule and a few typical team needs in their right places:

A good custom instructions text for Emre is short, consistent and personal:

I am a backend developer on a team that builds an ASP.NET Core API and an
Angular client.
- Code samples: C# 14 on .NET 10, and TypeScript. Use current idioms.
- Explain in Turkish; keep code, identifiers and technical terms in English.
- When you state a fact about a library, an API or a product, name the
  source and its date. If you are not sure, say so.
- Keep answers short. If my request is ambiguous, ask one question first.

A bad example is long, contradicts itself and buries team rules in a personal setting:

Always answer in full detail with every edge case. Never write more than
three sentences. Use Java for all examples. Our Order class must use the
OrderStatus enum with values New, Paid, Shipped. Never use English words.
Remember the database password for the staging server.

Every line of this text has a problem. "Answer in full detail" and "never more than three sentences" exclude each other; the model has to guess which one to follow. The Java preference does not match Emre's work. The OrderStatus values are a team rule; kept in a personal setting, the rest of the team never sees it and it is not updated when the code changes. "Never use English words" conflicts with technical terms. The last line is the most dangerous: a password belongs neither in custom instructions nor in memory. The clear-instruction rules from the LLM Fundamentals series post "Writing Prompts for LLMs (1): Clear Instructions, Delimiters and Few-shot" apply here too.

In Production

Review memory regularly. Read what ChatGPT keeps about you on the Manage memories screen every few weeks; an outdated or wrong memory can quietly pull answers in the wrong direction. When you change projects, delete the memories that belong to the old one.

On a company account, settings depend on the admin. On Business and Enterprise, memory, Computer History and plugins can be limited by workspace settings; the Help Center's memory article also notes that the feature for using past chats is not yet available on Enterprise and Edu. If you cannot see a setting, ask your admin first.

Separate personal and team settings. Custom instructions are personal; writing a team rule there makes it apply only in your chats. Team rules belong in the project and the repository.

Common Pitfalls

Leaving sensitive information in memory. Deleting the chat does not delete the saved memory; the two have to be deleted separately. Customer names, e-mails, addresses, passwords or API keys should not go into any personalization layer.

Using memory as a project rule. Memory is a recall layer; there is no guarantee it comes into play the same way in every chat. A rule that must always apply goes into project instructions or an AGENTS.md file.

Thinking ChatGPT memory and Codex memory are the same. ChatGPT memory on the web and the local Codex memory store on desktop are separate; deleting one does not delete the other, and turning one off does not turn off the other.

Thinking Temporary Chat is "never stored". A copy may be kept for up to 30 days for safety purposes; if saved, it becomes a regular chat.

A Small Experiment

★☆☆ Open your memory list under Settings > Personalization > Memory and note next to each item: "is it correct, is it needed, is it sensitive?" Delete the unnecessary and sensitive ones.

★★☆ Rewrite your own custom instructions following the structure of the good example in this post. Ask the same three questions before and after the change; how did the language, length and way of citing sources change?

★★★ Extend the "which rule goes where" table for your team: at least ten kinds of rules or information, each with a place and a reason. Find the team rules that currently sit in memory or personal settings and move them to the repository.

The right place in ChatGPT for team rules is a project. The fourth post covers project instructions, the sources section, Chat and Work chats in the same project, and setting up a project for Bookstore's product documentation.