How ChatGPT Memory Works and How to Manage It

Check which memory system appears in your account

Open Profile → Settings → Personalization → Memory. The labels matter because availability varies by plan, workspace, country, and rollout.

The current experience may show a Memory summary assembled from useful context in chats, files, and connected apps. OpenAI says the displayed summary may not include every detail that affects personalization. You can refresh it, correct an item, or ask ChatGPT not to mention something again.

Some accounts still show the older controls:

• Reference saved memories: retains details you explicitly ask ChatGPT to remember and may save useful conversational details.
• Reference chat history: lets ChatGPT draw on prior chats without representing every detail as an individual saved memory.

Follow the controls in your own account rather than an older screenshot. OpenAI maintains the ChatGPT Memory FAQ as the interface changes.

Memory, Custom Instructions, and Projects solve different problems

Custom Instructions

Use these for explicit, durable rules: your role, writing preferences, measurement units, programming language, or a response format you want consistently.

Memory

Use memory for personal context that emerges through conversation: a recurring project, a preference, or a goal that changes over time.

Projects

Use Projects for bounded work requiring related chats, source files, and project-specific instructions. Project instructions override global Custom Instructions inside that Project.

This separation makes errors easier to fix. If formatting is consistently wrong, change the instruction. If ChatGPT mentions an outdated priority, change the memory. For a broader comparison, see ChatGPT memory, Claude Projects, and Gemini.

Do not use memory or Custom Instructions as a document store. Keep long briefs, policies, and primary sources in files you control, then add them to the relevant Project. This reduces the AI context and memory bottleneck caused by repeatedly pasting the same material.

Review and remove what ChatGPT remembers

Start by asking, “What do you remember about me?” Treat the reply as a useful overview, not a complete audit.

Then open Memory settings and review anything that is:

• Stale
• Sensitive
• Vague
• Incorrect
• Likely to influence unrelated replies

Correct errors and remove context that no longer applies.

Deleting a chat does not necessarily remove a saved memory created from it. OpenAI says fully removing a detail can require deleting both the memory and the chat in which it was shared. The current system may also draw from files or connected apps, so review those sources when relevant.

Turning memory off does not delete information already stored. It stops that memory from being used while the setting is off. For privacy-focused guidance, see how to keep AI memory private.

Use Temporary Chat when a conversation should not access or create memories. It also stays outside regular chat history.

Use Projects for bounded work

A Project suits a client, course, job search, or long writing assignment. It groups chats, files, and project instructions.

When creating one, check its memory choice:

• Project-only memory: chats can use other conversations inside that Project but cannot use outside conversations or previously saved memories. OpenAI says this must be selected when the Project is created.
• Default memory: on personal plans, account memory can remain active in Project chats, and outside conversations may also be available depending on the plan. Project chats still prioritize their own chats and files.

Put local tone, format, and workflow rules in Project instructions instead of pasting them into every chat.

Save the right kind of context

A useful memory is short and durable: “I write Python for data work” or “I prefer vegetarian recipes.” A full brief that changes next week belongs in a file.

Review memory when a reply feels unexpectedly personal, outdated, or off-topic, and after major changes in role, location, health, or priorities. There is no universal audit schedule; audit when stored context stops matching your life.

If important material must work across assistants, keep the source outside any one model. Plain files, simple notes in Obsidian or Notion, a bookmark manager, or dEssence can hold information you need to find across tools. ChatGPT memory personalizes ChatGPT; it is not your independent archive.

Work around memory limits without guessing

Do not optimize around an unofficial word count. Behavior and management vary across accounts, and newer systems may prioritize or synthesize context instead of exposing a fixed list.

The reliable approach is to:

• Keep durable instructions explicit.
• Move long references into Projects.
• Maintain important source material elsewhere.
• Supply the source in the current chat or Project when exactness matters.

This is more dependable than trying to pack every fact into memory and provides a practical ChatGPT memory limit workaround.

Frequently asked questions

Is ChatGPT memory the same as Custom Instructions?

No. Memory carries conversational context into later replies. Custom Instructions are explicit rules you write.

Does deleting a chat delete its saved memory?

Not necessarily. Fully removing a detail may require deleting the saved memory and the originating chat.

What is project-only memory?

It limits a Project's chats to context from inside that Project, excluding outside conversations and previously saved memories. It must be selected when the Project is created.

When should I use Temporary Chat?

Use it for a one-off conversation that should neither access nor create memories and should not appear in regular chat history.

Where should exact facts and long briefs live?

Keep them in files you control and attach them to the relevant Project or current chat.