ChatGPT Slow or Out of Memory in a Long Chat? 3 Causes

When people type "ChatGPT out of memory" or "ChatGPT keeps running out of memory" into a search box, they are describing very different problems with the same phrase. Some are watching a browser tab grind to a halt as a conversation gets longer. Some have hit the message telling them the chat's memory is full. Some have noticed that their saved memories stopped updating weeks ago. All three get called running out of memory, and each one is fixed somewhere different.

This post sorts them apart. It starts with the one the other guides on this site do not cover, a long conversation making your browser or device struggle, and then deals briefly with the other two so you can tell quickly which one you have. The aim is to get you to the right fix in a few minutes rather than clearing settings that were never the problem.

Three Things That Get Called Out of Memory

Your device is running low. This is ordinary computer memory, the RAM your browser and operating system use. A very long conversation is a very long web page, and a long page with lots of formatted content takes more of your machine's memory to keep open and responsive. The symptoms are physical: typing lags, scrolling stutters, the tab goes blank or shows a crash page, the fan spins up, other tabs slow down too.

The conversation is full. This is the model's working budget for a single chat. Every reply is generated from the conversation so far, and there is a ceiling on how much of it can be carried at once. The symptoms are about the answers: the assistant forgets things from early on, contradicts settled decisions, or you get a message saying the chat's memory is full and offering a new chat with a summary.

The saved memory list is full. This is the separate account-level store of facts about you that new conversations draw on. When it reaches capacity, new facts simply stop being added. The symptoms are quiet: nothing errors, but new chats do not know things you told it recently.

The first is about your hardware. The second is about one conversation. The third is about your account. A fix for one does nothing for the others, which is why it pays to diagnose before acting. The split between the second and third is covered in depth in Context Window vs Memory; this post spends most of its time on the first.

A Two-Minute Diagnosis

Run these checks in order. They take less time than reading the rest of this post.

1. Is the whole tab sluggish, or just the answers? If typing into the message box lags, scrolling is jerky, or the page freezes, the problem is the page itself. Bad answers delivered on a responsive page point elsewhere.
2. Open a brand-new chat in the same browser. If the new chat is instantly smooth, the old conversation's length is what your device is struggling with. If the new chat is just as slow, the problem is broader: your browser, your machine, extensions, or the service itself.
3. Check other tabs and apps. If everything on the machine is slow, the long chat may be one heavy item among many rather than the whole cause.
4. Read any message on screen word for word. A message offering to start a new chat with a summary of this conversation is about the conversation budget, not your device. It is explained in detail in ChatGPT Chat Memory Full: What That Message Means.
5. Look at the newest entry in your saved memory list. If you use ChatGPT daily and the newest saved fact is from months ago, the list has probably stopped accepting new entries.

Most people get their answer at step one or two.

Why a Long Chat Can Strain Your Browser

None of the following depends on how any particular chat product is built internally. It is how web pages behave generally.

A conversation is rendered as a page of elements: each message, each code block, each table, each list. Formatting that makes answers readable, such as syntax highlighting and rich tables, also makes them heavier to display. A conversation with hundreds of exchanges, long pasted documents and a lot of generated code is a large page, and browsers need memory to keep large pages open and interactive.

Several things make it worse:

• Long pasted material. Pasting whole files, logs or documents into messages means the page holds all of it, often more than once if the assistant quotes it back.
• Lots of code. Highlighted code is among the heavier things to render, and coding sessions tend to run long.
• Many tabs of the same app. Each open conversation is its own page with its own memory use.
• Extensions. Any extension that reads or modifies the page does extra work on a long page. This includes extensions that add buttons to chat interfaces, MindLock's own included.
• Low-memory devices. An older laptop or a machine with many apps open has less headroom, so the same conversation that is fine on one computer is painful on another.

The model itself does not run in your browser in a normal chat product, so a heavy page is not the same as the model running out of room. The two often arrive together, simply because both get worse as a conversation gets longer, but they are different limits on different machines.

Fixes for a Slow or Crashing Tab

Start with the cheapest and stop as soon as things are usable.

Reload the page. A reload discards whatever has built up in the tab during a long session and renders the conversation fresh. Conversations saved to your history are still there after a reload; copy any long message you have typed but not yet sent before you do it. This alone often restores a usable tab for a while.

Close other tabs and heavy apps. Free memory is shared across everything on the machine. Closing a video call, a large spreadsheet or a dozen idle tabs gives the long chat room to breathe.

Try with extensions disabled. Open the conversation in a private or guest window where extensions are off by default, or temporarily disable them. If the tab is suddenly smooth, one of your extensions is adding load on long pages, and you can re-enable them one at a time to find which.

Update the browser. Performance and memory fixes ship in ordinary browser updates. If you have not restarted your browser in weeks, a restart that applies a pending update is worth trying.

Stop pasting whole documents. Paste the relevant section, not the whole file. This helps your browser and it helps the conversation budget too, since everything pasted is something the model has to carry.

Start a new chat with a handover. This is the real fix for a conversation that has simply become too long, and it solves the page problem and the conversation problem at the same time. A new chat is a small page and a fresh budget.

What does not help: clearing your saved memories, turning memory off, or deleting old conversations from your history. None of those change the size of the page you are looking at.

Writing a Handover That Survives the Move

The reason people put off starting a new chat is that they do not want to lose what the old one established. That fear is reasonable, and it is answered with five minutes of writing, not by staying in a conversation that is crashing your tab.

Before you leave the long chat, write down:

1. The goal in one or two sentences.
2. What is settled, one line per decision.
3. The constraints that must hold, one line each.
4. What was ruled out and why, one line each, because the reasons are what stop rejected ideas coming back.
5. The current state: where you stopped and what the next step is.

You can ask the assistant to draft this from the conversation, but read it before relying on it. A conversation long enough to be slowing your machine is often long enough that the model has already lost some of the early detail, so the draft is a starting point to correct rather than a finished record.

Paste the handover as the first message of a new chat. If you are mid-way through a technical task, add only the specific file or section you are about to work on, not the whole project. The technique of condensing a conversation yourself, and why it beats waiting for the product to do it, is laid out in How to Compact a Long ChatGPT Conversation Yourself.

Keep the old conversation in your history. It is still there to search and scroll back through if you need to check exactly what was said; it just should not be where you keep working.

When It Is the Conversation, Not the Device

If the page is responsive but the answers are going wrong, the device is fine and the conversation is the issue. The signs are forgetting early instructions, re-proposing ideas you rejected, losing track of names or numbers, or a message explicitly saying the chat's memory is full.

The fix is the same handover and new chat described above. The difference is urgency and expectation. A slow tab is uncomfortable but the answers may still be good. A conversation that has exceeded what the model can carry is producing answers from an incomplete picture, and every further message adds to the problem rather than fixing it. The longer you push on, the more you are working against a model that no longer has your early material in view. The full pattern, and how to spot it early, is in ChatGPT Loses Context Mid-Chat.

A useful habit: when you find yourself repeating an instruction you gave earlier, treat that as the signal to hand over. It almost always arrives before the explicit message does.

When It Is the Saved Memory List

If a new chat is smooth and fast, the answers inside conversations are fine, and the only problem is that ChatGPT does not seem to pick up new facts about you, look at the saved memory list in settings.

A full list does not announce itself loudly. New facts stop being added and that is all. The fix is pruning, not clearing: remove finished projects, superseded preferences and duplicates, and keep what is still true and useful. Copy the list somewhere first so a mistaken deletion can be put back by telling it the fact again. The order to prune in is set out in ChatGPT Memory Is Full: What to Prune.

This is the only one of the three problems that involves your settings at all, and it is the one most often wrongly treated as the cause of the other two.

Habits That Keep All Three in Check

Each of these problems is a growth problem. Things accumulate until they stop working. A few habits stop the accumulation before it bites.

Treat a conversation as a session, not a home. One chat per task or per working session, ended with a short handover. This keeps pages small, keeps the model's budget for the work that matters, and gives you a clean record of where things stand.

Paste sections, not whole files. Ask for what you need on this turn and paste only what that requires.

Restate constraints occasionally. In longer sessions, a one-line reminder of the constraints that must hold keeps them in recent context rather than buried near the start.

Review saved memories every few months. Ten minutes, asking of each entry whether you would want it applied to a conversation tomorrow.

Keep durable material outside the chat. Anything that has to survive the conversation ending, such as decisions, project briefs and preferences, should exist as text you own. Then a crashed tab or a full conversation is an inconvenience rather than a loss. Approaches to keeping copies of conversations you care about are in How to Back Up AI Conversations.

Where MindLock Fits

MindLock does not make long chats lighter for your browser and it does not raise the model's conversation limit. Those are properties of the chat product and of your device. What it addresses is the cost of leaving a conversation: keeping what mattered so that starting fresh does not mean starting from nothing.

You save a conversation worth keeping by pressing Ctrl or Cmd plus S in ChatGPT, Claude, Gemini or Perplexity and uploading the saved page on the Conversations page, as described in Importing Conversations. There is also a Chrome extension with a one-click Save to Memory button; it is loaded unpacked in developer mode rather than installed from a store, and like the keyboard route it saves only when you click. Distillation then turns saved conversations into memory documents, a profile memory plus topic memories, which you can search with Ctrl or Cmd plus K and use to generate a context block to paste into a new chat.

Two practical points. Saving and distilling are manual, so memory documents reflect conversations as of your last distillation rather than updating on their own. And free-tier distillation runs locally on your GPU through WebGPU, defaulting to Llama 3.2 3B, which needs a WebGPU-compatible GPU; it is a separate job from your chat, so it is best run when you are not also working in a heavy conversation tab. The Pro tier distils in the cloud instead, and the differences are on the pricing page.

The Short Version

Running out of memory in ChatGPT means one of three things. If the tab itself is slow or crashing, your device is struggling with a very long page: reload, close other tabs, test without extensions, and above all start a new chat with a handover. If the page is fine but the answers are forgetting things, the conversation has outgrown what the model can carry: hand over and start fresh. If new chats are fine but nothing new about you is being saved, the saved memory list is full: prune it in settings.

Diagnose first by opening a new chat and seeing whether it is smooth. Then fix the right thing. Clearing your memory settings is the right answer to only one of the three, and it is the one people reach for most often when they have one of the other two.