---
title: Why ChatGPT forgets, and how to tell which kind of forgetting you have
description: ChatGPT forgets for different reasons: a full context window, memory off, Temporary Chat, a project, a full memory list. Match your symptom to the cause.
date: 2026-11-18
updated: 2026-11-18
author: The Esment Team
category: Memory fixes
tags: chatgpt, memory, troubleshooting, context-engineering
canonical: https://esment.notas.ai/blog/why-chatgpt-forgets
---

# Why ChatGPT forgets, and how to tell which kind of forgetting you have

*The Esment Team · 2026-11-18 · [HTML version](https://esment.notas.ai/blog/why-chatgpt-forgets)*

## In short

- ChatGPT forgets in three different places: inside one long chat, between chats, and because of a setting or chat type. The symptom tells you which.
- Forgetting inside one chat is the context window. Early details fall out of view, and no setting changes that.
- A new chat that knows nothing about yesterday points to memory switches, Temporary Chat, a project's memory scope or a full memory list.
- Fix the cause you actually have. Restating a rule in a long chat does nothing for a Temporary Chat, and turning on memory does nothing for a full window.

ChatGPT forgets for six different reasons, and they have different fixes. **Start with the symptom.** If it loses details from earlier in the same conversation, you have hit the context window. If a brand-new chat knows nothing about yesterday, the cause is memory: a switch, a chat type, a project's scope or a full list.

This post is a diagnosis, not a settings checklist. For the step-by-step list of switches, see [ChatGPT memory not working? A checklist to find the cause](https://esment.notas.ai/blog/chatgpt-memory-not-working). For the concepts, see [how a context window works](https://esment.notas.ai/learn/the-context-window) and [how to make ChatGPT remember previous conversations](https://esment.notas.ai/learn/how-to-make-chatgpt-remember). Product details are as of November 2026, taken from OpenAI's help center, and some of them have changed more than once this year.

## Why ChatGPT forgets: three kinds of forgetting

People say "ChatGPT forgot" about three situations that have little in common.

- **In-chat:** something you said earlier in this conversation is gone. The cause is the context window.
- **Between chats:** a new conversation does not know what an old one established. The cause is memory, in one of several forms.
- **By design:** a setting or chat type is stopping memory from being read or written. Temporary Chat and project-only memory are the usual ones.

The rest of this post takes them in that order and ends with a short way to tell them apart.

## Reason 1: the context window fills up in a long chat

A model reads a conversation as one block of text, and that block has a size limit. When the conversation grows past it, the oldest material stops fitting, and the model works from what is left. That is why a rule you set in your first message can vanish an hour later. The [context window chapter](https://esment.notas.ai/learn/the-context-window) covers the mechanics.

**Figure.** A schematic window, not measured. As the chat grows, the early messages (highlighted) are the first to stop fitting.

- System + tools: 12%
- Saved memories: 6%
- Early messages: 20%
- Later messages: 50%
- Room to reply: 12%
OpenAI's [pricing page](https://openai.com/chatgpt/pricing/) says the usable window is smaller than the total, because system instructions, tools and memories (if enabled) take part of it. The sketch shows why: your conversation is never the only thing in there.

### How big is the window?

It depends on plan, model and mode, and OpenAI's pages do not fully agree. These are the non-reasoning figures from the [GPT-5.5 help article](https://help.openai.com/en/articles/11909943-gpt-5-5-in-chatgpt) and the pricing page:

| Plan | Standard (non-reasoning) window |
|---|---|
| Free | 16K tokens |
| Plus | 32K tokens |
| Pro | 128K tokens |

For the Thinking mode the sources conflict. A February 2026 [release note](https://help.openai.com/en/articles/6825453-chatgpt-release-notes) says 256K when you select it manually, the GPT-5.5 article says 256K for paid tiers and more for Pro, and the pricing page says 196K for several plans. We could not find confirmed window sizes for the newest GPT-6 models either. Treat every number here as "as published, check live".

The fix for this one is not a setting. Put the rules that matter in a short block at the top of a fresh chat, or start a new chat with a summary of where you stopped.

## Reason 2: a new chat starts empty

A new chat has an empty window. Nothing from yesterday is in it unless something puts it there. That something is Memory or chat history reference, and both are summaries rather than transcripts. OpenAI's [memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq) says the memory summary is a high-level synthesis, not a complete list.

So a new chat that knows "Alex Rivera is a product designer in Lisbon" but not the exact wording of last week's brief is behaving as designed. Saved memories hold facts and preferences. They do not hold your documents or your conversation. If you need a detail, bring it in yourself.

## Reason 3: memory or chat history reference is off

If a new chat knows nothing about you at all, check the two switches under **Settings › Personalization › Memory**. Labels vary by account, and may read Reference saved memories, Reference chat history or Memory summary.

The dependency is the thing people trip on. Per the memory FAQ, turning off Reference saved memories also turns off Reference chat history, while turning off chat history leaves your saved memories alone. The full walkthrough, including what happens to remembered data when you switch chat history off, is in the [memory checklist](https://esment.notas.ai/blog/chatgpt-memory-not-working), so we will not repeat it here.

## Reason 4: Temporary Chat and project-only memory

Two chat types forget on purpose, and both are easy to be in without noticing.

**Temporary Chat.** OpenAI's [Temporary Chat FAQ](https://help.openai.com/en/articles/8914046-temporary-chat-faq) says these chats do not use existing memories or create new ones. An August 2026 release note adds that a saved Temporary Chat becomes part of your chat history and can be referenced later if memory is on. Symptom: ChatGPT knows nothing about you, and nothing you say sticks.

**Project-only memory.** OpenAI's [Projects article](https://help.openai.com/en/articles/10169521-projects-in-chatgpt) says that with project-only memory, saved memories from outside the project are not referenced inside it, and chats see only other chats from the same project. The owner can change the setting in the project's settings, and changes may take a few hours. [How memory works in ChatGPT Projects and Claude Projects](https://esment.notas.ai/blog/chatgpt-claude-projects-memory) compares both products. Shared projects stay project-only. One forum report says the setting looks locked on some existing projects. That is a single user report, and we could not confirm it.

Symptom for both: ChatGPT knows you in ordinary chats and not here, or the reverse.

## Reason 5: the memory list is full or was reorganized

Saved memories have a capacity limit that OpenAI does not publish. When the list is full, new facts may not be saved. In October 2025, per the release notes, ChatGPT began managing saved memories automatically, keeping relevant details prominent and moving less important ones to the background, to help avoid that state. You can turn that off.

A June 2026 note says memories now update automatically and that the older saved-memories system can be restored from Settings › Memory › Saved memories. Rollout was described as still in progress, so what you see depends on your account.

Symptom: it used to remember something and no longer does, or a new fact never appears in the list. Our guide to [what ChatGPT's memory full message means](https://esment.notas.ai/learn/chatgpt-memory-full) explains the two limits and how to free space, so this post leaves the detail there. For a routine to keep the list tidy, see [a ChatGPT memory cleanup you can repeat](https://esment.notas.ai/blog/chatgpt-memory-full-cleanup).

> [!NOTE]
> "Chat memory full" or "Context window full" in the middle of a conversation is the window from Reason 1, not the saved-memory list. Deleting saved memories will not clear it.

## Reason 6: you changed model or uploaded a long file

Two more causes are worth a check when the first five do not fit.

**Switching model or mode.** Different models and modes have different windows, and Free and Go get different models than paid plans. So a switch mid-chat can leave you with a smaller window, and earlier material may stop fitting. This is our inference from the published limits, not something OpenAI states, so we say "can".

**Long uploaded files.** OpenAI's [file uploads FAQ](https://help.openai.com/en/articles/8555545-file-uploads-faq) lists upload caps, such as 512 MB per file and 2M tokens per text file. Those are limits on what you can upload, not on how much the model reads at once. We found no current OpenAI statement on how a long file is cut down, so do not assume the whole file is in view. Ask about named sections and check the details that matter.

## Diagnose it in 60 seconds

Match what you see to the cause. The table is the whole post in one place.

| What you see | Likely cause | Where to look |
|---|---|---|
| Early details of this chat are lost | Context window (Reason 1) | Start a fresh chat with a summary |
| New chat knows nothing about you | Memory or chat history reference off (3) | Settings › Personalization › Memory |
| New chat knows you but not last week's details | Memories are summaries, not transcripts (2) | Paste the detail in |
| Nothing you say is ever saved | Temporary Chat (4) | The label at the top of the chat |
| Knows you outside a project, not inside it | Project-only memory (4) | The project's settings |
| Used to remember, no longer does | Full or reorganized memory list (5) | Saved memories list |
| Lost track after switching model | Different window (6) | Switch back, or restart with a summary |
| Misses parts of a long document | File not fully in view (6) | Ask about specific sections |

> [!STEPS]
> 1. Ask: is the missing thing from this chat or from another one? This chat means the window; another chat means memory.
> 2. Check the top of the conversation for a Temporary label or a project name.
> 3. Open **Settings › Personalization › Memory** and confirm both switches.
> 4. Open your saved memories and see whether the fact is there at all.

## What helps, per cause

One short line each, since the full guides already exist.

- **Long chat:** restate the rules, or start fresh with a summary.
- **New chat:** check the switches, then ask ChatGPT to remember the fact by name. See [how to make ChatGPT remember](https://esment.notas.ai/learn/how-to-make-chatgpt-remember).
- **Temporary Chat or project:** use a normal chat, or change the project's scope.
- **Full list:** prune it, as in [the memory-full guide](https://esment.notas.ai/learn/chatgpt-memory-full).

## What a shared memory does and does not fix

ChatGPT's memory belongs to ChatGPT. A fact you told it does not reach Claude or Cursor, and the reverse is also true. If that is the limit you keep hitting, a memory kept outside any one assistant is a reasonable answer.

> [!ESMENT]
> Esment is one such memory. It keeps what matters in one place you can read and edit, and assistants you connect can search it. It does not enlarge ChatGPT's context window inside a single conversation, so Reason 1 stays Reason 1. You can [import your ChatGPT memory](https://esment.notas.ai/docs/memories/capture) to start, and [how Esment works](https://esment.notas.ai/learn/how-esment-works) explains the rest.

For why one assistant's memory is so easy to lose, read [why AI assistants need a memory they can't lose](https://esment.notas.ai/blog/why-ai-assistants-need-memory).

## Questions, answered

### Why does ChatGPT forget things in the middle of a long chat?

Every chat has a context window, a fixed amount of text the model can consider at once. When the conversation outgrows it, the oldest messages stop fitting and details drop out. Instructions, tools and memories also share that space, so you get less than the headline number.

### How big is ChatGPT's context window?

It depends on your plan, the model and the mode. OpenAI's pages we found list 16K, 32K and 128K for Free, Plus and Pro on non-reasoning chats, but figures for reasoning modes disagree between pages. Check OpenAI's pricing page for the current numbers.

### Why does a new chat not know what I said yesterday?

A new chat starts with an empty window. It only carries over what Memory or chat history reference supplies, and only if those are on. Saved memories are short notes, not transcripts, so the details of an old chat often do not appear.

### Does Temporary Chat make ChatGPT forget?

Yes, by design. OpenAI says Temporary Chats do not use existing memories or create new ones. A release note from August 2026 adds that a saved Temporary Chat joins your chat history and can be referenced later if memory is on.

### Can switching models make ChatGPT forget?

It can. Different models and modes have different context windows, and some are not offered on Free or Go. If you switch mid-chat, a smaller window may not hold as much history. That is our inference from the published limits, not an OpenAI statement.

### Does ChatGPT read my whole uploaded file?

Not necessarily. Upload caps, such as 2M tokens per text file, are limits on what you can upload, not on what the model reads at once. We found no current OpenAI statement on how long files are cut, so ask about specific sections and check key details.
