ChatGPT memory limit: is there one, and what is it?

There is. Three, in fact, and people mean different ones. Here is what OpenAI publishes for each, plan by plan, as read on 7 September 2026 and re-checked on 15 September; the one capacity figure it has published, and the absolute numbers it never has; the few levers that genuinely widen what ChatGPT remembers; and why the ceiling that costs you most at work isn't a number at all.

Short answer: ChatGPT's memory is bounded, and OpenAI has said so twice on the same day. Its 4 June 2026 announcement describes efficiency work that made it possible "to increase memory capacity for Plus and Pro users." Its ChatGPT release note dated 4 June 2026 puts a figure on that increase: "For Plus and Pro users, ChatGPT can also remember more useful context, with twice as much memory capacity." That multiplier is the only capacity figure OpenAI publishes anywhere, and it is relative — twice their previous capacity, whose size OpenAI has never published — on two plans out of four. The absolute size of the bound is unpublished, on every plan. So a specific count of memories quoted as "the ChatGPT memory limit" isn't OpenAI's, and neither is a multiplier for Free or Go. The numbers OpenAI does publish in absolute terms belong to a different limit, the context window, and those it sets plan by plan.

Three limits, one search term

"ChatGPT memory limit" gets used for three separate ceilings. Working out which one you're hitting is most of the answer.

The memory limit itself: real, and unpublished

Start with what memory is, in OpenAI's words. When enabled, it "helps ChatGPT automatically remember useful context from your chats, files, and connected apps to personalize your experience, so you don't have to repeat yourself as often." On the current system that context is held as a memory summary — "a continually updated synthesis of context from your past chats," with ChatGPT "keeping track of the details it determines are most important." The older saved memories list, where each remembered fact was a separate entry, is still reachable from Settings → Memory via the "saved memories" link beneath the summary, and OpenAI now calls it the legacy system.

Now what OpenAI says about the size of either. Read in full on 7 September 2026, the Memory FAQ contains no entry count, no capacity figure and no "memory is full" state — for the summary or for the legacy list. That silence isn't a denial, and it also isn't the whole public record, because the release notes are a third surface and they carry two things the FAQ doesn't. First, a figure: the entry dated 4 June 2026 says Plus and Pro accounts get "twice as much memory capacity" — relative, never absolute, so it sizes the change and not the ceiling, and it says nothing about Free or Go. Second, a state: the entry dated 15 October 2025 says automatic memory management "helps prevent saved memories from reaching capacity and helps avoid ‘memory full’ state in your ChatGPT account." Read that one narrowly — it describes the legacy saved-memories list, it says "helps avoid" rather than prevents, and the same entry notes you "can turn off automatic memory management at any time," which hands the pruning back to you. Nothing equivalent is documented for the memory summary. One number does get misquoted as a memory figure, so it's worth being precise: the "approximately 5x" in the June announcement describes a reduction in the compute needed to serve memory to Free users. It says nothing about how much memory anyone gets.

The dated sequence, from the announcement: memory launched in April 2024 as saved memories; in April 2025 ChatGPT gained the ability to reference chat context beyond that list, through a background process OpenAI calls dreaming; on 4 June 2026 the current architecture — labelled Dreaming V3 in OpenAI's own evaluations — went to Plus and Pro accounts in the US first, "and will roll out to additional countries and Free and Go users over the coming weeks."

Two practical consequences follow. OpenAI documents no indicator of how close you are to the bound — the FAQ describes the summary and the per-answer sources view, and nothing that reports fullness for it. The legacy list is the one place OpenAI's own wording implies a full state you'd be told about. And the summary page isn't the store: OpenAI says it "will not include everything that ChatGPT remembers based on your chats," and suggests that if you want to know whether something was remembered, you "just ask in chat."

ChatGPT memory limit: Free vs Go vs Plus vs Pro

The plan comparison is the one place OpenAI ranks memory by tier, and it does so in words. From the ChatGPT pricing page, read 7 September 2026:

So "free vs Plus" resolves to Limited versus Expanded, and Plus versus Pro to Expanded versus Expanded in the comparison table — though the Pro plan card says "Maximum." OpenAI doesn't reconcile the two labels; our reading of its wording is that Pro is at least Plus, and the page says nothing more precise than that. No plan is listed as unlimited, and no plan shows a number. If a page tells you the memory limit on Plus is a specific count, it isn't quoting OpenAI.

Memory versus the context window, in OpenAI's words

This is the conflation behind most "memory limit" complaints, and OpenAI defines the two separately. Memory, from the FAQ, is above. The context window, from a footnote on the pricing page: "ChatGPT manages a shared context window to understand your request, track the conversation, retrieve relevant information, and generate responses. The portion available for user input is smaller than the total window, as space is also used for system instructions (including tools and personality), memories (if enabled), and internal processing (reviewing information, reasoning, and response generation)."

Unlike memory, the window has published sizes. From the same page, same day:

Read that footnote once more, because it settles a common misunderstanding: memories, when enabled, take up part of the window. A fuller memory doesn't give a conversation more room. OpenAI's own footnote says the space for your input "may change dynamically based on features in use and any memory content." So "make ChatGPT remember more" and "stop ChatGPT losing the brief halfway through a long chat" are different problems with different fixes. The second one has its own guide. The rest of this page is about the first.

How to increase ChatGPT's memory: what actually moves it

OpenAI documents no dial. What it does document, in descending order of effect:

• The plan. The only documented lever on the bound itself. Free is "Limited," Go "Yes," Plus and Pro "Expanded," and the June update doubled the Plus and Pro bound, whose size OpenAI has never published. That's the whole of what changes with money.
• Move fixed facts into Custom Instructions. OpenAI's own line: "For explicit information or instructions, you can add it to your Custom Instructions. For information shared via conversations, ChatGPT can remember relevant details for you." Your role, your clients' names, how you want answers formatted — those don't need to compete for room in a synthesis that decides what's "most important."
• Move working material into a Project. Projects are available on every plan, you can create as many as you like, and files attached to one persist for every chat inside it — 5 files per project on Free, 25 on Go and Plus, 40 on Pro. For Plus and Pro accounts OpenAI says ChatGPT "prioritizes the project chats and files" when you ask a question there, and a project can be set to project-only memory (Project settings → Memory) so one client's context stays out of another's. One caveat from OpenAI's own FAQ on projects: project memory "does not show a list of memories like personal memory" — you can't audit what it kept.
• Steer the summary instead of re-explaining. Type what you want changed into the text box at the bottom of the memory summary, or highlight a passage to correct it. The summary's menu has a Refresh option if it looks stale. On the legacy list, "Remember that…" adds an entry directly, and OpenAI says ChatGPT can update, combine or remove saved memories when asked.
• Keep the noise out. Temporary Chats "do not use existing memories or create new memories," so a throwaway session never spends whatever room the bound gives you.

And what doesn't work: the pricing page lists no memory add-on you can buy on top of a plan, and "clear it to make room" isn't a documented fix for the summary — OpenAI's own explanations for an empty summary are a new account, recently enabled memory, or too little chat history, not fullness. If details have genuinely stopped sticking, the controls and their limits are in what to do when ChatGPT memory is full.

What to keep outside ChatGPT's memory

Here is the part the plan table can't tell you. Whatever the bound is, memory is a synthesis — a running profile of you, built from "the details it determines are most important." It was never designed to hold the client's Tuesday email verbatim, or the redline three replies deep, or what the CFO actually said about pricing in the October review. OpenAI's route to that kind of material is files and connected apps: upload it by hand, into a chat or a project, or connect a whole account so memory can draw on it.

That route has a cost on the way out, too. To fully delete something ChatGPT may know about you, OpenAI says, "you'll need to delete every source where it appears, including past chats, archived chats, files, the memory summary, and disconnect any connected apps that may contain that information." The more you feed it to widen the memory, the more places a single detail lives.

So the working rule is simple. The memory summary is for who you are and how you like things done. The material behind an engagement belongs either in a project, inside ChatGPT, or somewhere built to hold the originals and hand them back — outside it.

The limit underneath: source, not size

Suppose the bound vanished tomorrow. You'd still be choosing between two shapes of memory: a summary of you assembled from what passed through ChatGPT, or a record of the work — the thread itself, the PDF itself, the note you dictated in the car — kept where a question can reach it and the original is one tap away. For anyone running seven client relationships out of an inbox, the second shape is the one that saves a meeting, and no memory tier turns the first into it.

That's the shape Equerry is built around, and the gesture is deliberately small: CC your assistant. It reads the thread — not your inbox. Add your Equerry address to the client thread, or forward the one PDF that matters; drop in a screenshot from the share sheet; dictate a voice memo after the call; save a note. Nothing gets in unless you send it — there's no inbox connection to grant. Ask "what did Acme agree on pricing?" and the answer comes with a source chip pointing at the exact email or note it came from, so you open the original rather than trust a summary. Before the call, a morning brief pulls together what you've shared. And in Settings, every fact it has kept is listed and deletable on a tap, per note; delete the note and what came from it goes too. It's a personal assistant that remembers what you handed it and shows its work — available on the App Store for iPhone and iPad.

Keep ChatGPT for what it's good at, and use the levers above to get the most from its memory. Just don't expect a bigger bound to solve a problem that was never about size. For what ChatGPT actually carries between chats, see does ChatGPT remember past conversations; for how an assistant built around your own sources compares, see a ChatGPT memory alternative for work context.

Every OpenAI statement above was read on 7 September 2026 from four OpenAI pages, with a fifth added and the capacity statements re-verified on 15 September 2026: the Memory FAQ (help.openai.com, article 8590148) for how memory works, the memory summary, sources, Temporary Chats and the legacy saved-memories system; the 4 June 2026 announcement "Dreaming: Better memory for a more helpful ChatGPT" (openai.com) for the timeline, the rollout and the single capacity sentence; the ChatGPT — Release Notes article (help.openai.com, article 6825453) for the entries dated 4 June 2026 (the "twice as much memory capacity" figure) and 15 October 2025 (automatic memory management and the legacy "memory full" state); the ChatGPT pricing page (chatgpt.com/pricing) for the per-plan memory wording, the context-window sizes and the shared-window footnote; and the Projects in ChatGPT article (help.openai.com, article 10169521) for file limits and project-only memory. help.openai.com refuses automated fetches with a 403, so its articles were read through a text-extraction reader proxy and confirmed separately in a browser session, both on 15 September 2026. Quoted wording is OpenAI's. Where those pages are silent — the absolute size of the memory bound on any plan, and what a full memory summary does — this page says so rather than filling the gap. OpenAI changes these pages often; treat your own Settings screen as the authority for your account.

Last updated 2026-09-15.

All product names, logos, and brands are property of their respective owners. Equerry is not affiliated with, endorsed by, or sponsored by any other product mentioned; comparisons are provided for informational purposes.