ChatGPT Memory

What is ChatGPT Memory?

ChatGPT Memory is OpenAI's personalization system that allows ChatGPT to retain information about a user and apply it across otherwise separate conversations. Without memory, every chat starts blank and ChatGPT only knows what is in the current context window. With memory enabled, details such as your name, job, preferences, and recurring projects can persist so you do not have to repeat them each session. OpenAI first introduced the feature for ChatGPT in February 2024 and has expanded it several times since.

Memory is built from two distinct layers. The first is saved memories: discrete facts ChatGPT has recorded, which you can view, add to, and delete individually under Settings. The second is reference chat history (controlled alongside a setting OpenAI labels "Reference saved memories"), in which ChatGPT draws on the broader content of your past conversations rather than a fixed list of facts. Together these let ChatGPT personalize answers using both explicit stored items and a synthesized sense of your history.

As of mid-2026, the underlying memory is produced by a background process OpenAI calls Dreaming. The latest version, Dreaming V3, began rolling out on June 4, 2026. It reads across many past conversations asynchronously and maintains a synthesized memory state that is injected into the model's context at the start of each new chat, so a fresh conversation can already reflect what the system has learned about you.

• Personalization layer that carries information across separate ChatGPT conversations.
• Two layers: editable saved memories plus reference chat history (cross-chat recall).
• First launched in February 2024; expanded through 2025 and 2026.
• As of mid-2026, memory is generated by OpenAI's background "Dreaming" synthesis process.
• Distinct from the single-session context window, which resets each new chat.

Saved memories vs reference chat history: the two layers

Saved memories are the explicit layer. These are individual items ChatGPT has chosen to store, such as "prefers concise answers" or "is learning Spanish." You can ask ChatGPT to remember something directly, ask what it remembers, and review or delete entries in Settings. When ChatGPT writes a new saved memory during a chat, it typically shows a brief "Memory updated" notice, giving you a visible record of what was added.

Reference chat history is the broader layer. Rather than relying only on the fixed list of saved memories, ChatGPT can draw on relevant information from your earlier conversations to inform new responses. OpenAI's own example is that if you once said you like Thai food, it may take that into account the next time you ask what to have for lunch. This layer is not a clean, item-by-item list, so OpenAI's help documentation notes that the memory summary you can see does not necessarily include everything ChatGPT may remember from your history.

The two layers are linked in the settings. According to OpenAI's Memory FAQ, turning off "Reference saved memories" also turns off "Reference chat history," while you can keep saved memories on and turn reference chat history off on its own. OpenAI also states that turning off reference chat history deletes the information ChatGPT had learned from past chats, with that data removed from its systems within 30 days.

• Saved memories: discrete, editable items you can view, add, and delete individually.
• Reference chat history: broader recall synthesized from your past conversations.
• ChatGPT usually shows "Memory updated" when it writes a new saved memory.
• Turning off "Reference saved memories" also turns off reference chat history.
• Turning off reference chat history deletes learned data within about 30 days.

What changed with Dreaming V3 (June 2026)?

Dreaming is OpenAI's name for the background synthesis process that builds memory from your conversations without explicit prompting. Dreaming V3 began rolling out on June 4, 2026. Instead of depending on a manually curated list of saved facts, a single asynchronous background process reads across many past conversations at once, captures context that comes up naturally, and updates existing memories as your circumstances change. OpenAI's example is a note that you are going to Singapore in July rewriting itself to reflect that the trip has happened, so a past event is not later referenced as if it is still upcoming.

The synthesized memory state is maintained in a separate data layer and injected into the system prompt at inference time. In practice, this means every new conversation starts with a context that already contains what the system has synthesized about you, rather than ChatGPT looking things up only when triggered. OpenAI describes Dreaming as having supplemented saved memories over the prior year to produce what it calls a step-function improvement in personalization.

OpenAI reports internal evaluation gains, citing factual recall rising from roughly 67.9 percent in its earlier system to about 82.8 percent with Dreaming V3, alongside preference adherence of 71.3 percent and time-sensitive accuracy of 75.1 percent. These figures are OpenAI's own and, as press coverage has noted, have not been independently audited. OpenAI attributes the wider rollout to a roughly 5x reduction in the compute cost of running the dreaming process, which let it extend memory to Free-tier users, reaching Plus and Pro subscribers in the United States first with Free, Go, and international users following.

A practical consequence is that the visible saved-memories list no longer fully represents what ChatGPT knows. Because memory is synthesized in the background, deleting a single conversation does not by itself remove memories derived from it. OpenAI states that logs of deleted saved memories may be retained for up to 30 days for safety and debugging.

• Dreaming V3 rollout began June 4, 2026, as a background memory-synthesis update.
• Reads many past chats asynchronously and updates memories as circumstances change.
• Synthesized memory is injected into the system prompt at the start of each chat.
• OpenAI cites factual recall improving from ~67.9% to ~82.8% (its own, unaudited figures).
• A ~5x compute reduction enabled the rollout to Free-tier users.

How to view, manage, and turn off ChatGPT Memory

You manage memory under your profile picture, then Settings, then Personalization. There you can see whether memory is on, open the controls for saved memories and reference chat history, and review the individual memories ChatGPT has stored. From the manage-memory view you can delete a specific memory or clear all saved memories at once. You can also act conversationally: tell ChatGPT to remember or forget something, or ask it what it currently remembers about you.

To stop ChatGPT from recording new memories, turn off the memory toggles in Personalization. Remember that turning off "Reference saved memories" also disables "Reference chat history," and that disabling reference chat history deletes the data ChatGPT learned from past chats, with removal from OpenAI's systems within about 30 days. Deleting individual saved memories removes those entries, though OpenAI notes deletion logs may persist for up to 30 days.

For a one-off private conversation, use Temporary Chat. A Temporary Chat does not use your saved memories or chat history and is not used to update memory, so nothing from it should persist or influence future personalization. This is the recommended path for sensitive topics you do not want folded into your long-term memory.

• Manage memory under Profile, then Settings, then Personalization.
• Delete a single memory or clear all saved memories from the manage-memory view.
• Disabling reference chat history deletes learned data within about 30 days.
• Use Temporary Chat for sessions that should not read from or write to memory.
• You can also tell ChatGPT to remember, forget, or list what it remembers.

Memory vs custom instructions: how they differ

Custom instructions and memory both personalize ChatGPT, but they are managed differently. Custom instructions are explicit, static settings you write yourself, telling ChatGPT what you want it to know about you and how you want it to respond. They stay exactly as you typed them until you change them, which makes them predictable and a good fit for stable preferences such as tone, format, or role.

Memory, by contrast, is updated automatically by ChatGPT based on a continually refreshed synthesis of your past chats. OpenAI's help documentation describes saved memories as working similarly to custom instructions, except the models update them automatically rather than requiring you to manage each item by hand. Because memory is synthesized in the background, it can be broader than the list of items shown in your memory summary.

A useful rule of thumb: put explicit, durable instructions in custom instructions, and let memory capture the details that surface naturally across conversations. Memory is also separate from the context window, which is only the text in the current conversation and resets when you start a new chat. Tools that act as a dedicated, user-controlled memory layer across apps, such as MemX, take a different approach again. MemX is private by architecture, with per-user isolation and encryption at rest, rather than relying on a single vendor's account-bound personalization.

• Custom instructions: explicit, static, written and managed by you.
• Memory: automatically synthesized and updated by ChatGPT from past chats.
• Memory can be broader than the visible list of saved-memory items.
• Both are distinct from the context window, which resets each new conversation.
• Use custom instructions for durable rules; let memory handle organic details.