How ChatGPT Memory Works

An assistant that learns to know you

Without memory, every conversation starts from scratch: you have to restate that you’re an accountant, that you prefer short answers, or the name of your company. With Memory enabled, ChatGPT can retain preferences and useful details from your interactions, using them to personalize future responses. Understanding what it remembers, where it pulls that information from, and what it leaves out lets you get the most out of the feature while keeping control of your data.

Context, memory, instructions: three different things

Context is what ChatGPT sees in the current conversation; it disappears when you start a new one. Memory persists across conversations and evolves over time. Custom instructions, finally, are guidelines you write yourself that apply to all your conversations; Lesson 7 compares these to memory.

Where memory gets its information

Depending on your plan and region, memory can draw from several sources: your past conversations, saved memories, your custom instructions, files stored in your Library or Space, and content from connected apps like Gmail. It doesn’t keep every detail of every conversation: ChatGPT decides what’s relevant for a given response, and what it retains evolves with your situation.

Two mechanisms coexist. Saved memories are pieces of information you explicitly ask it to remember, or that it records on its own as useful context when this feature is available:

Remember that I work at a 50-person SME
specializing in digital transformation consulting.

Chat history reference additionally allows ChatGPT to pull from your past conversations what it deems useful; this information changes over time as it updates what deserves to be remembered. In the enhanced memory experience, ChatGPT maintains an overall summary of this context, which you can read and correct.

See what was used

When ChatGPT relies on your personal context, a Sources mention may appear below the response. It indicates what contributed to the personalization: a past conversation, a memory, an instruction, a file, or an email. Depending on the case, you can open the source, correct or delete a memory, or flag that a source wasn’t relevant. These Sources don’t necessarily show every factor, and they aren’t included when you share a conversation via link.

What memory doesn’t do

Three scenarios fall outside memory’s scope. A temporary chat creates or modifies no memories, even if you opened it in personalized mode. In a locked conversation, new messages aren’t used to personalize your regular chats. And when memory is disabled, nothing new is retained. Memory capacity also depends on your plan: limited on Free, longer on Go, extended on Plus, and maximum on Pro.

The effect on your conversations

With memory active, ChatGPT adapts its responses to your profile without you having to repeat your context. If you’ve mentioned being a Python developer, its code examples will be in Python. If you prefer short answers, it will get straight to the point. If you’re working on a specific project, it will refer to it: you ask “draft the invitation email,” and it already knows which event you mean. The gain is subtle but cumulative. The downside is too: false or outdated information in memory silently skews future responses, which is why Lesson 6 exists.

Common misunderstandings

Believing ChatGPT remembers everything: memory retains what it deems useful, not the entirety of your exchanges. Never checking what’s retained: an error in memory propagates quietly. Storing sensitive data there: never ask ChatGPT to remember a password, login credential, or confidential information. Confusing memory with context: in a long conversation, it’s the thread’s context that matters; memory primarily serves future conversations.

Your turn

Your memory in three steps. Open Settings > Personalization > Memory and note the settings available for your account (Memory, Reference saved memories, Reference chat history, memory summary…). In a new conversation, say: “Remember that my company is called TechNova and that we have 30 employees.” Open another conversation and ask: “Write a three-line presentation of my company.” Look for the Sources mention below the response, if there is one, and verify where the information came from.

Official sources: Memory in ChatGPT (OpenAI Help Center: memory sources, saved memories, history, Sources, temporary and locked chats) · ChatGPT Pricing (memory by plan)

Key takeaways

• Memory persists between conversations; context stops at the current conversation
• It can draw from your past conversations, saved memories, your instructions, your files, and certain connected apps
• A Sources mention below the response shows what was used and allows for corrections
• Temporary chat: no memories created; locked conversation: no personalization of other threads
• Never entrust passwords or confidential information to memory