Don't stop here
Hand-picked guides our readers explore right after this one.
Master ChatGPT with advanced prompting techniques, mega-prompts, and proven frameworks
Read the guideExpert guide to Claude prompts with XML tags, artifacts, and complex reasoning
Read the guideBuild a compelling resume with AI assistance
Read the guideGPTPrompts.AI Editorial
Tested across Free, Plus, and Team workspaces on web, iOS, and macOS in May 2026 · Last updated May 15, 2026
Quick answer
ChatGPT Memory has two parts in 2026: saved memories (an explicit, editable list) and reference chat history (implicit recall). Both live at Settings, Personalization, Memory. Each can be toggled off. Temporary Chat bypasses both. Memory syncs across iOS, Android, macOS, Windows, and web for one OpenAI account.
Below: a feature comparison vs Custom Instructions and Projects, a tier-by-tier availability table, the full rollout timeline with OpenAI announcement dates, and the prompts we use to audit and curate what ChatGPT has saved about us.
How we tested this
We compared what ChatGPT can and cannot remember across a Free account, two Plus accounts (one US, one EEA), and a Team workspace, running the same prompts on web, the iOS app, and the macOS desktop app.
For each behavior described below, we checked Settings, Personalization, Memory to confirm the toggle state, used Temporary Chat as a baseline for 'no memory', and reviewed OpenAI's public announcements and help-center docs to source every date and tier claim. Where OpenAI's public docs were the canonical source, we cite the announcement date.
In 2026, ChatGPT's cross-chat continuity is actually two distinct systems. The settings UI lists them as two separate toggles, and they fail differently if you turn one off.
An explicit list of facts ChatGPT has chosen to remember. Each entry is editable and deletable. Visible at Settings, Personalization, Memory, Manage. ChatGPT confirms a new save inline with 'Memory updated'.
Good for
Implicit recall of patterns from past chats, launched April 10, 2025 (per OpenAI's announcement on @OpenAI). No list view; ChatGPT just draws on past chats when relevant.
Good for
Key difference: Saved memories are auditable; reference chat history is not. If you want to know exactly what ChatGPT thinks it knows about you, only the saved-memories list is observable. The chat-history layer is inferred at runtime from past conversations and never displayed as a fixed list.
In our testing
The most useful prompt we landed on was the audit prompt: 'List everything you currently remember about me. Format as bullet points. Group by category (work, preferences, projects, personal). Be exact, do not paraphrase or smooth over.' Running that monthly was the single highest-ROI Memory habit we adopted; it caught two stale facts that were quietly shaping responses we did not want.
The chat-history reference layer was both a delight and a small problem. We loved that ChatGPT could pick up a project from a chat three weeks ago without re-pasting context. We disliked that it occasionally pattern-matched into the wrong project (mentioning a client codename from a different chat thread). Saying 'Forget anything from the chat where I asked about X' did not always work; the reliable fix was to delete the source chats themselves.
Temporary Chat became our reflex for personal-life questions. Memory is fantastic for ongoing work, less fantastic when it remembers you researched a private health question once. Use the Temporary Chat icon (next to the model picker) for anything you do not want shaping ChatGPT's view of you. We tested that Temporary Chat was bypassed end to end, no saved memory was written and the chat did not appear in history, exactly as documented.
On Team workspaces, the surprise was how clean the separation was. Memory on our Team account stayed strictly within the workspace. The same person signing into a personal ChatGPT account saw their own memories, not the Team ones. That is the correct behavior per OpenAI's Team docs, but worth verifying yourself before you put anything sensitive in a memory.
Three overlapping ChatGPT features for persistence. Use the right one for the job.
| Feature | Memory | Custom Instructions | Projects |
|---|---|---|---|
| Persists across chats | Yes (saved memories) and Yes (reference chat history) | Yes (applies to every new chat) | Yes, but only within the same Project |
| User can see and edit each entry | Yes for saved memories. Reference chat history is implicit, not a list. | Yes, free-text 'What should ChatGPT know about you' fields | Yes, per-Project instructions and files |
| Available on Free tier | Yes (smaller saved-memory cap; chat-history reference rolled out 2024) | Yes | Yes (Projects went broadly available in 2024) |
| Available on Plus and Pro | Yes, with larger saved-memory cap | Yes | Yes |
| Available on Team and Enterprise | Yes (Team default on; Enterprise off-by-default, admin-controlled per OpenAI Enterprise docs) | Yes | Yes |
| Survives Temporary Chat | No, Temporary Chats are excluded from memory by design | No, Temporary Chat bypasses Custom Instructions | Not applicable (Projects use their own chats) |
| User-controllable wipe | Yes (Settings, Personalization, Memory, Manage / Clear all) | Yes (clear the text fields) | Yes (delete the Project) |
| Used to train future models by default | No on Team / Enterprise / Edu by default. On Free / Plus / Pro, controlled by 'Improve the model for everyone' toggle in Data Controls. | Same control as memory | Same control as memory |
Rule of thumb: Custom Instructions for a stable identity (always speak in this tone, always include code in TypeScript). Memory for evolving context (a project that lasts months). Projects for tightly scoped contexts that should not leak into other chats (client work, an investigation, a thesis).
Both Memory layers are widely available, but caps and default states differ by tier.
| Tier | Saved memories | Reference chat history | Default | Notes |
|---|---|---|---|---|
| Free | Limited count visible in Settings, Personalization, Memory (smaller than Plus) | Yes, rolled out broadly 2024 | On | Free users can disable both toggles independently. Memory does not unlock unlimited storage. |
| Plus ($20/month) | Larger saved-memory cap than Free | Yes | On | When the saved-memory cap fills, ChatGPT prompts you to delete older entries before adding new ones. |
| Pro ($200/month) | Same UX as Plus; reasoning models use the same memory store | Yes | On | GPT-5 Thinking, o3, and o3-pro share the same saved memories and chat-history reference as the default conversational model. |
| Team ($25/user/month annual; $30 monthly) | Yes, scoped to each member's own workspace login | Yes | On | Memory respects workspace data-control settings; Team data is not used to train OpenAI models by default per OpenAI's Team docs. |
| Enterprise | Yes, admin-controlled | Yes, admin-controlled | Off (admin must opt in) | Enterprise admins enable or disable memory at the workspace level. Enterprise data is not used to train OpenAI models per OpenAI's Enterprise privacy commitments. |
| Edu | Same as Enterprise; admin-controlled | Same as Enterprise; admin-controlled | Off (admin must opt in) | ChatGPT Edu mirrors Enterprise data-handling commitments per OpenAI's Edu launch announcement. |
Five dates explain how Memory grew from a small test to the two-layer system that exists in 2026. Each is sourced from OpenAI's own announcements.
Rolled out to a small percentage of Free and Plus users for testing.
OpenAI confirmed the regional carve-out via the X account @OpenAI on launch day.
Team defaulted on; Enterprise required admin opt-in.
Announced on @OpenAI on X; arrived first to Plus and Pro, with Free, Team, and Enterprise following per OpenAI's product rollout pages.
EEA / UK rollout staged later than the US per OpenAI's regional rollout notes; check Settings, Personalization, Memory to confirm the toggle exists in your account.
Three independent surfaces affect what ChatGPT remembers about you. Knowing all three is the difference between thinking you turned Memory off and actually turning it off.
Settings, Personalization, Memory. Two toggles: 'Reference saved memories' and 'Reference chat history'. Disable both for no continuity. Toggles are independent.
Settings, Personalization, Memory, Manage, Clear all. This deletes every saved-memory entry. Past chats are NOT touched; ChatGPT will only forget what is in the saved-memory list.
Settings, Data Controls. 'Improve the model for everyone' toggle controls training-data use on Free, Plus, Pro. 'Delete all chats' wipes past conversations so the chat-history layer has nothing to draw on.
Full reset recipe: Step 1: Settings, Personalization, Memory, toggle off both layers. Step 2: Click Manage, Clear all. Step 3: Settings, Data Controls, Delete all chats. Step 4: For a really clean start, sign out and sign back in. After this, only Custom Instructions and any active Projects shape future conversations.
Our verdict
Use Memory if you use ChatGPT daily for ongoing projects (a multi-month book, a long-running codebase, a thesis). Saved memories plus chat-history reference cut prompt-setup time meaningfully, often a minute or two per session. Audit monthly with the list-everything prompt.
Use Custom Instructions instead if the thing you want ChatGPT to know is stable (your role, preferred tone, default coding language). Static fields beat dynamic memory for stable identity because they cannot be silently overwritten by an inline 'remember that'.
Use Projects instead if the work belongs to a tight scope you do not want bleeding into other chats (client work, sensitive research, a thesis). Projects keep memory and files contained, with their own per-Project instructions.
Use Temporary Chat (NOT Memory) if the question is personal, sensitive, or one-off. Anything you research in Temporary Chat is excluded from Memory by design. Reserve regular chats for work you want ChatGPT learning from.
Disable Memory entirely if you share an OpenAI account, work in a regulated environment where every saved fact is a small audit liability, or simply prefer ChatGPT to start fresh every session. Disabling both toggles costs nothing on the answer-quality side; Custom Instructions still applies.
Our overall take: Memory is a net positive for daily users once you adopt the monthly audit habit. The two failure modes (stale facts shaping responses, occasional cross-project pattern matching) are mitigated by curating the saved-memory list and using Temporary Chat for off-topic queries. The chat-history reference layer is more useful than the saved-memory layer for most people; saved memories are most useful for hard constraints (allergies, accessibility, tooling defaults) that you do not want ChatGPT to forget.
The questions readers ask most about ChatGPT Memory.
The line we draw after a year of daily use. Putting these into Memory creates more risk than convenience.
Memory is not a vault. Even when training-data use is off on Team / Enterprise, the value lives in an LLM-readable system. Use a password manager.
Future answers will be subtly shaped by what Memory thinks it knows. Health context belongs in a chat scoped to that task, not your permanent profile.
Names, addresses, and personal details of friends, family, or coworkers. They did not consent to being in your saved memories.
If you researched something once because you were curious, do not let it shape every future response. Use Temporary Chat next time.
Same logic as passwords. Memory is convenience, not security. The risk dwarfs the time-saving.
Keep reading
Master ChatGPT with advanced prompting techniques, mega-prompts, and proven frameworks
Read guide →Create stunning images with ChatGPT DALL-E integration, viral photo trends, and professional visuals
Read guide →Master Midjourney from v4 to v6 with expert techniques
Read guide →ChatGPT prompts for coding interview prep covering algorithms, system design, and behavioral questions
Read guide →AI prompts for Excel formulas, macros, data analysis, automation, and dashboards
Read guide →Write, debug, and explain code with ChatGPT
Read guide →