Memory & Cognition

Reference Chat History

Reference chat history is the ChatGPT memory setting that lets the model draw on information from your past conversations automatically, without you asking it to remember anything. It is the implicit cross-chat layer that sits alongside the editable "Saved memories" list and can be toggled on or off under Settings > Personalization > Memory.

What is Reference Chat History?

Reference chat history is a ChatGPT memory setting that allows the model to use information from your earlier conversations to personalize future responses, without you explicitly asking it to store anything. It is one of two memory mechanisms ChatGPT exposes: the other is "Saved memories", a list of discrete facts you ask it to keep. OpenAI's help documentation describes the behavior this way: if you once mentioned that you like Thai food, ChatGPT may take that into account the next time you ask what to have for lunch. The setting lives under Settings > Personalization > Memory, where it appears next to the toggle labeled "Reference saved memories".

The feature is sometimes called the "chat history" half of ChatGPT memory, in contrast to the "saved memories" half. Where saved memories are explicit and editable, reference chat history is implicit: ChatGPT decides which details from across your conversations are worth carrying forward, and that set can shift over time. OpenAI's wording is that details from past chats can change as ChatGPT updates what is more helpful to remember, which is why the company advises using saved memories for anything you want it to always retain.

OpenAI first announced that ChatGPT could reference past conversations on April 10, 2025, expanding memory beyond the original saved-memories list that launched in early 2024. As of mid-2026 the underlying machinery has been rebuilt around a background synthesis process OpenAI calls Dreaming, but the user-facing toggle name has stayed the same. "Reference chat history" remains the exact string most people search when they want to find, understand, or disable the setting.

It is the implicit, cross-chat memory layer in ChatGPT, distinct from the editable Saved memories list.
When on, relevant context from past conversations can be added to new ones automatically.
It appears in Settings > Personalization > Memory beside "Reference saved memories".
OpenAI introduced past-conversation referencing on April 10, 2025.
The remembered set is not fixed; it changes as ChatGPT updates what is useful to keep.

How does Reference Chat History differ from Saved Memories?

Saved memories are details you explicitly ask ChatGPT to remember, such as your name, your dietary preferences, or how you like answers formatted. They appear as individual entries you can read, edit, or delete, and according to OpenAI they are considered in future responses until you remove them. Reference chat history works differently: nothing is added by you on purpose, and there is no fixed list of entries that map one-to-one onto specific things you said.

The practical distinction is permanence and visibility. A saved memory is durable and inspectable; a detail surfaced through reference chat history is best-effort and may fade or be revised as ChatGPT reorganizes what it finds useful. OpenAI's guidance is explicit that ChatGPT does not retain every detail from past chats, so anything you need it to always know should be a saved memory rather than something you hope it picked up from conversation.

The two settings are also linked in how they can be switched off. Per OpenAI's help docs, turning "Reference saved memories" off also turns "Reference chat history" off, but if "Reference saved memories" is on you can still turn "Reference chat history" off independently. In other words saved memories is the parent control and chat history is the dependent one.

Saved memories: explicit, editable, durable until you delete them.
Reference chat history: implicit, not a one-to-one list, and subject to change over time.
Use saved memories for anything you want ChatGPT to remember reliably.
Disabling "Reference saved memories" also disables "Reference chat history".
You can disable chat history alone while keeping saved memories active.

What changed with the Dreaming V3 synthesis?

In 2026 OpenAI rebuilt the chat-history layer around a background process it calls Dreaming, with the latest iteration referred to as Dreaming V3. Rather than reacting to individual conversations one at a time, an asynchronous process reads across many past chats together and synthesizes a higher-level model of the user, updating that model as circumstances change. OpenAI began rolling out Dreaming V3 on June 4, 2026, starting with Plus and Pro users in the United States before extending to Free, Go, and international users in the following weeks.

A defining property of the new design is that synthesized memory is time-aware. OpenAI's own example is that a note like "You are going to Singapore in July" can be revised to "You went to Singapore in July 2026" once the trip has passed, so stale context gets corrected instead of accumulating. OpenAI reported large benchmark gains for the synthesis approach, including factual recall rising to roughly 82.8 percent and time-sensitive accuracy to about 75.1 percent, alongside preference adherence near 71.3 percent, though these are vendor-reported figures rather than independent measurements.

Architecturally, reporting on the rollout describes the synthesized memory state as living in a separate data layer that is injected into the system prompt at inference time, rather than being stored inside the visible conversation log. One consequence flagged by coverage of the update is a visibility tradeoff: the memory summary surface does not necessarily list everything ChatGPT may carry forward, so the older sense of a fully auditable memory list is weaker under Dreaming than it was with explicit saved memories alone.

Dreaming is the background synthesis engine behind modern reference chat history.
Dreaming V3 began rolling out June 4, 2026, US Plus and Pro first.
Synthesized memory is time-aware and revises outdated context automatically.
OpenAI reports factual recall near 82.8 percent (vendor figure, not independent).
The synthesized state is injected at inference time, not stored in the chat log.

How do you view, manage, or turn off Reference Chat History?

To change the setting, open ChatGPT, go to Settings, then Personalization, and find the Memory controls. There you will see "Reference saved memories" and "Reference chat history" as separate toggles. Switching "Reference chat history" off stops ChatGPT from pulling context out of your past conversations into new ones, while leaving any explicit saved memories in place if that parent toggle is still on.

Managing what is remembered is more nuanced than a single switch. Saved memories can be opened and deleted individually from the same settings area. Chat-history-derived context is harder to enumerate, and coverage of the Dreaming update notes an important caveat: deleting a conversation does not by itself remove memories that were derived from it, so removing a detail can require clearing both the conversation and the relevant memory. Reporting also indicates that logs of deleted saved memories may be retained for up to 30 days for safety and debugging.

If you want a conversation that ignores memory entirely, ChatGPT offers Temporary Chat, which does not use or update saved memories or chat history and is not saved to your history. Temporary Chat is the cleanest way to get a one-off, memory-free session without changing your global settings. For a permanent change, the Personalization toggles are the right place; for a single session, Temporary Chat is the better tool.

Path: Settings > Personalization > Memory, then toggle "Reference chat history" off.
Saved memories can be reviewed and deleted individually in the same panel.
Deleting a chat does not automatically erase memory derived from it.
Temporary Chat ignores both saved memories and chat history for a single session.
Disabling the setting stops future personalization but does not retroactively scrub everything.

Why does Reference Chat History matter for privacy and memory control?

Reference chat history changes what "deleting a chat" means. Because the synthesized memory layer is separate from the raw conversation log, a deleted conversation can still have contributed to what ChatGPT carries forward about you. For anyone who treats chat deletion as a privacy action, the practical takeaway is that controlling the memory setting and reviewing the memory summary matter as much as clearing history, and that the summary may not be exhaustive under the Dreaming architecture.

It also raises ordinary questions about scope. Reference chat history operates per account and is tied to your settings; it is not designed to share what you said in one account with other users, and OpenAI exposes per-user controls to turn it off. The reasonable posture is to treat it as a convenience feature you opt into, verify the toggle state, and use Temporary Chat or a disabled setting when you do not want a conversation feeding future personalization.

Tools built specifically around memory control take a different design stance. A memory layer that is private by architecture, for example, can keep per-user isolation, encryption at rest, customer-managed keys, and on-device options, so the user decides what is stored and retained rather than inheriting an implicit synthesis. That is a design choice about control and inspectability, not a claim that any single ChatGPT setting is unsafe; the point is that reference chat history is implicit by default, so knowing where its controls live is the main lever you have.

Synthesized memory is separate from the chat log, so deleting a chat is not a full erase.
The memory summary may not list everything the model can carry forward.
The setting is per account with user-level controls to turn it off.
Temporary Chat or a disabled toggle keeps a session out of future personalization.
Memory-control tools that are private by architecture emphasize per-user isolation and explicit retention.

Key takeaways

Reference chat history is ChatGPT's implicit cross-chat memory setting, distinct from the explicit, editable Saved memories list.
OpenAI introduced past-conversation referencing on April 10, 2025, and the toggle lives in Settings > Personalization > Memory.
Turning off "Reference saved memories" also turns off "Reference chat history", but you can disable chat history alone.
In 2026 the layer was rebuilt around Dreaming (Dreaming V3 began rolling out June 4, 2026), which synthesizes time-aware memory across many chats.
Deleting a conversation does not automatically remove memory derived from it, so chat deletion is not a complete erase.
Temporary Chat ignores both saved memories and chat history, making it the simplest memory-free session option.

Frequently asked questions

It is the memory setting that lets ChatGPT use information from your past conversations to personalize new ones, without you asking it to remember anything. It is separate from Saved memories, which are explicit entries you create. OpenAI describes it as recalling useful details, like a food preference you mentioned earlier, to make later chats more relevant.

Go to Settings > Personalization > Memory, and toggle "Reference chat history" off. Note that turning off "Reference saved memories" also turns off chat history, but you can switch chat history off on its own while keeping saved memories on. For a single memory-free session, use Temporary Chat instead.

Saved memories are explicit, editable entries kept until you delete them, while reference chat history is implicit context ChatGPT pulls from past chats and can revise over time. OpenAI does not retain every detail from past chats, so anything you want it to always know should be a saved memory. Saved memories is also the parent control that, when off, disables chat history.

It operates per account with user-level controls, and you can turn it off in Settings. However, because the synthesized memory layer is separate from your conversation log, deleting a chat does not automatically remove memory derived from it, and the memory summary may not list everything. To keep a session out of personalization, disable the setting or use Temporary Chat.

OpenAI began extending memory that references recent conversations to free users on June 3, 2025, with paid tiers getting longer-term, higher-capacity memory. As of mid-2026, the Dreaming V3 rollout started with US Plus and Pro users before expanding to Free, Go, and international users in the following weeks.

Dreaming is OpenAI's background synthesis process that builds a model of the user from many past chats at once, with Dreaming V3 rolling out June 4, 2026. It makes memory time-aware, so outdated context like a past trip date gets revised rather than kept stale. The synthesized state is injected at inference time and is not stored in the visible chat log.

Put the idea into practice

MemX is an AI memory app built on these ideas: store anything, skip the folders, and find it again by asking in plain English.

Try MemX Free