Memory, custom instructions, and Projects

You told ChatGPT last month that you manage a five-person ops team, prefer bullet lists, and hate vague “synergy” advice. This morning it drafts a customer email in the wrong voice, forgets your product name spelling, and invents a meeting cadence you never agreed to. You did not break the product. You just treated three different “memory” features as if they were one sticky brain.

This post is part of a series that walks through ChatGPT from the very beginning. Earlier posts covered what ChatGPT actually is, how to pick a plan without hype, and the first thirty minutes on web and mobile. They also covered how the desktop app splits into Chat, Work, and Codex. This one maps out custom instructions, memory, Projects, and, on some paid plans, scheduled tasks, so you can see how each one actually works. Get this part right and you stop pasting the same company brief into every new chat. You also stop leaving sticky personal details in places you never meant to keep them.

Product labels move, so treat OpenAI’s own Help Center articles on custom instructions, memory, Projects, scheduled tasks, and data controls as the source of truth, and open those pages again before you write team policy. The menu paths below match the Help Center as it stood in September 2026: Settings → Personalization on web and desktop, and Customize ChatGPT on mobile for some toggles.

Three boxes, not one brain

People say “ChatGPT remembers me” the way they say “the laptop knows my Wi-Fi.” What’s really going on is that you are managing several separate stores of information. Each one has its own rules, its own place where you go to edit it, and its own way of going wrong.

Picture three boxes side by side, plus a fourth clock for tasks:

• Custom instructions. Text you write on purpose. It applies across chats once you turn customization on. This is the right place for standing style and role, such as “I’m a data analyst; prefer tables; skip the fluffy opener.”
• Memory. Context the product keeps so you do not have to re-explain yourself every session. You can review it, edit it, and delete it. Free plans tend to feel more limited in how much sticks and how often it gets used, so do not store sensitive facts here by accident.
• Projects. A workspace that groups chats, reference files, and instructions around one ongoing goal. A project can keep its own memory focused on that project’s history instead of pulling in everything else you have ever typed.
• Scheduled tasks (some paid plans). One-off or repeating prompts that ChatGPT runs later and then tells you about. They work well for reminders and light monitoring, but they are not a real job scheduler and they are not your CRM’s workflow engine.

Rule of thumb: Custom instructions are the laminated card on your desk. Memory is the sticky notes the product keeps rewriting. Projects are the labeled binders. Scheduled tasks are the calendar invites. Treat them as four separate tools, not one magic recall.

Custom instructions: standing prefs you control.

Custom instructions are for anyone tired of pasting the same preamble into every chat. OpenAI’s Help Center describes them as guidance ChatGPT should consider in its answers, applied whenever customization is turned on. You write them and you edit them, so they will not quietly invent new “facts” about your company from a late-night chat you barely remember having.

Where they live (typical paths)

• Web and desktop: Settings → Personalization → Custom Instructions. Turn on customization, then write the text.
• iOS and Android: Settings → Customize ChatGPT. Toggle customization on, then type in the instructions.

Updates apply to future answers right away, but old chat transcripts still show whatever the model said back then. If a past thread is stuck in the wrong voice, delete or archive that chat instead of assuming a new instruction rewrites the history.

Character limits (plan-dependent)

As of September 2026, OpenAI’s Help Center puts the limit at roughly 1,500 characters for Free and Go accounts, and up to 5,000 characters for Plus, Pro, Enterprise, Business, and Education. Treat those numbers as facts you should recheck later, not as something fixed forever. The practical advice does not change either way: shorter instructions that stay true beat a long block that quietly contradicts itself two paragraphs down.

What to put in (and what to leave out).

Good global instructions read like a short style guide, not a vault of secrets:

• Role and audience (“I write for non-technical managers; define jargon once”).
• Output shape (“lead with the answer; then bullets; then risks”).
• Hard “don’ts” that are not secrets (“no emojis; no fake citations; flag uncertainty”).
• Light durable prefs (“US English; metric units unless I say otherwise”).

Leave out Social Security numbers, passwords, API keys, card numbers, private customer lists, raw HR files, and anything your employer treats as restricted. Instructions are convenient, not a secure vault. OpenAI also notes that if you use third-party plugins or tools, relevant parts of your instructions may be shared with those tools, so only connect integrations you actually trust.

Example instruction block you can adapt.

Role: I am an analytics lead at a mid-size B2B company.
Audience: busy managers who want clear numbers and plain English.
Format: start with the answer in 2 sentences, then bullets, then risks or unknowns.
Style: concrete examples; no hype; no fake studies; say when you are guessing.
Defaults: US English; prefer tables for comparisons; ask one clarifying question if the metric grain is unclear.
Do not: invent tool screenshots, invent legal advice, or invent SQL that I did not ask to verify.

That block fits inside the Free-tier character limit if you trim a line or two, and it is specific enough that every new chat starts less random. For tone that only applies to one piece of work, such as a marketing launch versus your personal tax notes, use project instructions instead of stuffing everything into the global field.

Memory: sticky context you should audit.

Memory is the feature people mean when they say “it finally stopped forgetting my job title.” When memory is on, ChatGPT holds on to useful context from your chats. That way, personalization carries over into next week instead of resetting every time. OpenAI’s Memory FAQ is the page to reread whenever the interface shifts between “saved memories,” “memory summary,” and related controls, because those names change more often than the underlying feature.

Editable is the point

You are not stuck with whatever the model inferred about you. Typical controls, which you can confirm under Settings → Personalization → Memory, include:

• Turn memory on or off entirely.
• Review a memory summary or a saved memory list.
• Edit or delete specific items.
• Ask in chat what it remembers, then correct it in plain language.
• Use Temporary Chat when you want a one-off conversation that should not create or use memories.

Deleting a chat does not always erase every memory that chat produced. Treat memory cleanup as its own small chore, the way you clean out browser cookies after a long research binge. If something sensitive showed up in memory, remove it from memory and, if you still care about the residue, from the chat history too.

Free plan limits (practical view)

Memory can feel thinner on Free, and that is not a sign you are using the tool wrong. It is a product tier choice: Free is built for trying the product and light daily use, not for building a second brain out of your entire career. Paid consumer plans usually give you more room and more personalization features. Exact caps change over time. If Free starts to feel forgetful, either upgrade for volume you will actually use, or keep durable facts in Project files. Do not expect the general app memory to store everything forever.

Memory vs custom instructions

OpenAI’s own guidance is simple: put explicit standing rules in custom instructions, and let memory carry the conversational details you would rather not retype. When the two disagree, fix the source of truth instead of arguing with the model in five different chats.

What not to let memory keep.

• Credentials, one-time codes, private keys.
• Customer data dumps, full HR packets, or medical details you would not put in a shared Google Doc.
• Unverified gossip about coworkers.
• Temporary project codenames that must not leak into personal chats.
• Anything your company policy marks as off-limits for consumer AI tools.

If you need a place for messy drafts, Temporary Chat, where it is available, is built so those chats are not saved to your history the normal way and do not create memories. OpenAI also says Temporary Chats are not used for training in the usual consumer sense described in its Data Controls help page. Check the Temporary Chat FAQ for the current retention window; OpenAI describes keeping a safety copy for about 30 days to watch for abuse.

Projects: binders for long-running work

A long sidebar full of chats named “Untitled” is how people lose track of decisions. Projects solve that organizing problem first: one place for the chats, files, and instructions that belong to a single goal. OpenAI documents Projects for both free and paid accounts. Business, Enterprise, and Edu plans get richer team sharing and admin controls. Consumer plans can share projects too, with collaborator and file limits that vary by plan. Check the current numbers in Help.

What a Project holds

• Chats created inside the project, or moved in from elsewhere.
• Files you upload as reference, such as a PDF, a spreadsheet, a doc, an image, or pasted text, with limits that depend on your plan.
• Project instructions that apply only inside that project and override your global custom instructions.
• Project memory so context stays focused on that project, including project-only options useful for sensitive or long-running work.
• Optional saved responses kept as project sources when a draft is worth reusing.

Project instructions override global ones

This is the part people miss. Your global custom instructions still apply to everyday chat outside a project. Inside a project, the project’s own instructions win instead. That is how you keep a calm, concise default for personal use while a brand-voice, checklist-heavy style runs inside “Q3 launch brief,” without rewriting the global field every Monday.

Project: Q3 launch brief
Instructions:
- Act like a careful product marketing partner.
- Always ground claims in the uploaded brief and competitive notes.
- Flag anything not in the files as a guess.
- Prefer short headlines, then bullets, then open questions.
- Never invent pricing or ship dates.

Project-only memory vs default memory

When you create a project, you may be able to choose project-only memory instead of the default setting. Project-only means chats in that project lean on context from inside the project and do not pull in your broader personal memories the way a regular chat would. Shared projects are usually forced into this project-only boundary, so a teammate does not accidentally inherit details from your other, private chats. If you do sensitive work, prefer project-only and keep your personal memories out of that binder.

File limits change, so plan for a cap.

OpenAI’s Help Center, checked in September 2026, lists file counts per plan: around 5 files on Free, and up to 25 or 40 on higher tiers. Storage and upload rate limits also apply across chats and Projects. When you hit the wall, delete stale files or split the work into more than one project. You can also keep the master document in a company drive and paste only the sections you need. The next post in this series covers that paste-and-upload decision in more depth.

Project starter kit (20 minutes)

1. Name it by goal, not by tool, so “Weekly ops review” instead of “ChatGPT stuff.”.
2. Add one brief file with goals, audience, definitions, and non-goals. One honest page beats ten half-related PDFs.
3. Add one or two example outputs you already like, such as a past email or a report section, because models copy shape better than they copy vibes.
4. Write 3 to 8 lines of project instructions covering role, format, and hard don’ts.
5. Chat only inside that project for that goal, for at least a week, and move old related chats in if the interface allows it.
6. Save strong answers back into project sources when the product offers a “Save to project” option.
7. Review memory and files monthly. Delete the junk and update the brief once reality changes.

OpenAI’s own examples include a school study guide built from PDFs and study mode, a marketing launch plan, and personal event planning. For AMS readers, strong candidates are “metric definition cleanup,” “customer interview synthesis,” and “SQL review notes for team onboarding,” each with a short glossary file and two good sample answers already attached.

Scheduled tasks: paid automation, not magic ops.

Scheduled tasks let you ask ChatGPT to run a prompt later, whether that is a one-time reminder, a recurring checklist, or a monitoring-style check that only pings you when something meaningful changes. OpenAI documents them for Plus, Pro, Business, and Enterprise, and both availability and the number of active tasks vary by plan. If you are on Free, assume you do not get the full task system until your account or the Help Center says otherwise.

What they are good for

• Personal reminders with a useful draft attached, such as a Sunday-evening prompt to draft your weekly priorities from a checklist.
• Light recurring research digests on public topics.
• Monitoring-style prompts, on products that support “notify only on change”.

What they are bad for

• Mission-critical monitoring where a missed alert has a real cost; use a real observability tool instead.
• Anything that needs guaranteed access to private project files, since OpenAI notes that a task created inside a project may not be able to see that project’s files.
• Voice chats and custom-GPT-driven tasks, where documented gaps exist, so check Help if you depend on that combination.
• Secrets written into the task instructions, because that prompt will run unattended later.

Create tasks from the Scheduled page in the sidebar on web, mobile, or the desktop Chat experience, or by asking for one inside a chat. Edit, pause, and delete them from the task interface. OpenAI’s Help Center, checked in September 2026, put active task limits around 5 for Plus and higher for higher tiers, which is a good reason to prune zombie reminders the same way you prune calendar spam.

Worked example: one personal stream, one work stream.

Say you use ChatGPT on a Plus plan. You write your global custom instructions once: plain English, tables for comparisons, never invent citations. You turn memory on for convenience. Then every Friday you open Memory and delete anything sensitive, such as passwords, customer lists, or old job search notes.

From there you create two Projects:

When you need a throwaway “is this joke funny?” chat, you use Temporary Chat so it does not pollute memory. When you want a Monday-morning nudge to list open tickets from a summary you will paste in again, you set up a scheduled task with no credentials written into the prompt. When your company later moves you onto Enterprise, you read the admin policy again, because workspace training defaults and retention can look very different from your old consumer account.

Privacy and training: consumer opt-out vs work defaults.

Personalization features sit right next to data controls. The two are related, but they are not the same setting.

• Consumer plans (Free, Plus, Pro, and similar): you can usually opt out of training under Data Controls by turning off Improve the model for everyone. Your chat history can still exist after you do this; training is a separate question from storage. Check the Data Controls FAQ again for the exact signed-in and signed-out paths.
• Business, Enterprise, and Edu: OpenAI documents stronger default privacy for these plans, including not training on business workspace content by default. Admins control the feature toggles, retention, and sharing, so your personal habits do not automatically become company policy.
• Projects training note (consumer): Free, Plus, and Pro project content may be used for training when “improve the model” is on. Shared projects carry extra rules about when training applies across contributors, so read the Projects Help section on data, not just the marketing page.

None of this turns a consumer chat into a locked medical-records vault. If your employer issues a work account, use it for work. If you only have a personal account, keep work data off it unless your company’s policy specifically allows that tool for that kind of data.

Common mistakes with instructions, memory, and Projects.

Do this in Settings today

1. Open Settings → Personalization. Write or trim your custom instructions to under one screen.
2. Open Memory. Delete anything you would not paste into a shared team doc.
3. Create one Project named for a real goal. Add a one-page brief and two example outputs.
4. Run three chats inside that Project and nowhere else. Save one good answer as a source if the interface allows it.
5. If your plan includes tasks, create one low-stakes recurring reminder with no secrets in the prompt, then pause it once you confirm the notification works.
6. Open Data Controls and decide, on purpose, whether Improve the model for everyone should be on. Match that choice to how sensitive your everyday prompts actually are.

Quick recap

• Custom instructions are the standing prefs you write, and you can turn them off any time.
• Memory is sticky context you have to edit and prune yourself, and it is more limited on Free.
• Projects scope your chats, files, and instructions to one goal, with their own project memory options.
• Scheduled tasks live on some paid tiers; good for reminders and light monitoring, not critical operations.
• The consumer training opt-out lives in Data Controls; Enterprise privacy is a different contract entirely.
• Never store credentials or regulated data “so it remembers” them.

Series notes

This is Part 5 of the Learn ChatGPT from scratch series, on personalization. Earlier posts covered what ChatGPT is, how to pick a plan, the first thirty minutes on web and mobile, and the desktop Chat, Work, and Codex split. The next post covers voice, images, and files for everyday tasks. That includes the paste-and-upload ladder: never passwords or account numbers, high caution with customer or HR data, care with internal docs, and free room to explore with public data. Later in the series, another post covers workplace rules and when not to use ChatGPT at all. If you are comparing this to agent tools, the Claude Code memory layers post on this site is a related way of thinking about it, not a substitute for these ChatGPT settings.

Sources

Vendor help pages checked in September 2026. Re-open the live page before you treat a character limit or file cap as fixed forever.

• OpenAI Help: ChatGPT Custom Instructions (enable paths, character limits by plan, data notes).
• OpenAI Help: Memory FAQ (how memory works, editing, Temporary Chat contrast, instructions vs memory).
• OpenAI Help: What is Memory (capability overview entry point; cross-link from product docs).
• OpenAI Help: Projects in ChatGPT (files, project instructions, project-only memory, sharing, plan file limits, training notes).
• OpenAI Help: Scheduled Tasks in ChatGPT (who can use tasks, caps, monitoring tasks, unsupported combos).
• OpenAI Help: Data Controls FAQ (Improve the model for everyone, Temporary Chats, export pointers).
• OpenAI Help: Temporary Chat FAQ (no history/memories/training behavior as documented).
• OpenAI Help: ChatGPT Capabilities Overview (memory, Projects, scheduled tasks in one map).
• OpenAI: Enterprise privacy (how business workspace data is treated vs consumer defaults).
• Analytics Made Simple: Learn (related learning paths on this site).

Merch on Etsy

Open store →

Naive Bayes Coffee Mug View on Etsy → Mean Squared Error Shirt View on Etsy → You're On Mute Coffee Mug View on Etsy → Standard Deviation White Ceramic Mug View on Etsy →