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 guideUnlock Google's Gemini with multimodal prompting strategies
Read the guideMaster Midjourney from v4 to v6 with expert techniques
Read the guideYou have memory switched on, you have talked to ChatGPT about your project for weeks, and it still opens every new chat as if it has never met you. The confusion here is that memory is not one feature. There are two independent systems with two separate toggles. Saved memories are an explicit, auditable list of facts, and you can read and edit every entry. Chat history reference is an implicit layer that draws on patterns from your past conversations, and you cannot inspect what it holds or when it will surface. Most 'ChatGPT does not remember' reports come down to the second toggle being off, or to expecting the implicit layer to behave like a database it can be queried against. It does not: recall is probabilistic, it is stronger for stable facts and preferences than for specific details from one conversation, and it will not reliably retrieve what you said in a particular chat last Tuesday. This page explains which system you are actually hitting, how to check both, and the reliable workaround when you need context to carry across sessions with certainty.
New chats start with no awareness of ongoing work discussed elsewhere
ChatGPT recalls general preferences but not specifics from a particular conversation
It remembers some things and not others with no obvious pattern
Nothing is recalled at all despite the memory toggle being on
Recall works on the web but not in the mobile app, or the other way around
It remembers something you asked it to forget months ago
Saved memories and chat history reference are two different switches under Settings, Personalization, Memory. Having saved memories on does nothing for cross-chat recall, which is the other toggle. This single misunderstanding accounts for most reports.
Chat history reference surfaces patterns and durable context, not a lookup of past transcripts. It is good at your role, your stack, your writing preferences, and unreliable at 'what did I decide about the pricing table three weeks ago'. Expecting the latter is expecting the wrong feature.
Saved memories are added when the model judges something durable, or when you ask directly. Casual mentions in passing frequently are not stored at all, so there is nothing to recall later even with the feature on.
Temporary Chats deliberately do not read or write memory. Projects scope context to the project, so work inside one may not surface in general chats and vice versa. Both behave exactly like memory being broken if you do not know which mode you are in.
Memory is per account. A different login, or a work account under an enterprise or education plan where memory may be disabled by an administrator, produces a complete blank with no error message.
When to try: First, before anything else
Go to Settings, Personalization, Memory and confirm that saved memories and reference chat history are both enabled. They are independent. Turning on only the first is the single most common reason cross-chat recall does not work.
When to try: When recall is partial or inconsistent
In the same settings panel, open the list of saved memories. This shows exactly what ChatGPT holds as explicit facts. If the thing you expected it to know is not in the list, it was never saved and no amount of prompting will retrieve it.
When to try: For any context you need to persist reliably
Say 'Remember that...' followed by the fact, and check afterwards that it appears in the saved memories list. Confirmation in the chat is not proof it was stored. Explicit saves are auditable and durable, which is why they are the right home for anything that matters.
When to try: When memory appears completely absent
Temporary Chats do not use or update memory by design, and they look identical to normal chats apart from the indicator. If you habitually open them, that alone explains total memory failure. Start a standard chat instead.
When to try: For any ongoing multi-session piece of work
Put related conversations in a Project and attach the relevant files and instructions. Project context applies consistently to every chat inside it, which is far more dependable than hoping the implicit layer surfaces the right thing at the right moment.
When to try: For preferences that should apply every single time
Role, stack, tone, and formatting preferences belong in custom instructions rather than memory. Custom instructions are applied to every chat deterministically, while memory recall is probabilistic. This is the reliability upgrade most people are looking for when they complain about memory.
When to try: When correctness across sessions genuinely matters
Maintain a short document with your project state, decisions, and constraints. Attach it or paste it at the start of a session. It is manual, and it is the only method that guarantees the model has the exact context you intend, which matters when the cost of a wrong assumption is high.
When to try: When memory works in one place and not another
Check the account in the profile menu, especially if you use both personal and work logins. Memory does not cross accounts, and on managed workspaces an administrator may have memory disabled entirely, which presents as memory simply never working with no explanation shown.
Keep both memory toggles on and review your saved memories list occasionally
Put stable preferences in custom instructions, not memory, so they apply every time
Use Projects to group work that needs shared context across many sessions
Maintain a short project context file you can attach whenever precision matters
Contact OpenAI support through help.openai.com if both memory toggles are on, entries are visible in your saved memories list, and none of them are ever reflected in new chats across multiple devices and a fresh browser. Also report memories that reappear after deletion, since that is a genuine defect rather than expected recall behavior. Include your plan, the platform, whether the failure is web-only or app-only, and screenshots of the memory settings panel. If you are on a work or school account, check with your administrator first, because memory can be turned off at the workspace level with no visible notice to you.
Cross-chat recall is a separate setting from saved memories. Under Settings, Personalization, Memory there are two toggles, and having saved memories on does nothing for referencing past conversations. Turn on the chat history reference toggle as well. If it is already on, note that recall is pattern-based rather than a transcript lookup, so specific details from one past chat may still not surface.
Saved memories are an explicit list of facts you can open, read, edit, and delete in settings. Chat history reference is implicit: it draws on patterns from your past conversations, you cannot see what it contains, and you cannot control when it applies. Anything you need reliably should be an explicit saved memory or a custom instruction, not left to the implicit layer.
Say 'Remember that...' and then verify the entry appears in the saved memories list in settings, because a confirmation in the chat does not prove it was stored. For preferences that should apply every time, custom instructions are more reliable than memory because they are applied deterministically to every conversation.
No. Temporary Chats do not read from or write to memory by design, and they are not added to your history. If your normal habit is opening Temporary Chats, memory will appear completely broken even when it is configured correctly. Check the indicator at the top of the conversation.
The implicit layer favors stable, repeated context (your role, your tools, how you like answers formatted) over one-off specifics mentioned once in passing. It is also curated automatically in the background rather than storing everything, so recall genuinely is uneven. Move anything that matters into a saved memory, a custom instruction, or a Project.
Product behavior and limits can change. These primary sources were used to verify this guide.