---
title: How memory works in ChatGPT Projects and Claude Projects
description: How memory works in ChatGPT Projects and Claude Projects: instructions and files vs memory, project-only memory, what carries between chats, and setup for each.
date: 2026-11-18
updated: 2026-11-18
author: The Esment Team
category: Guides
tags: chatgpt, claude, memory, projects
canonical: https://esment.notas.ai/blog/chatgpt-claude-projects-memory
---

# How memory works in ChatGPT Projects and Claude Projects

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

## In short

- A project holds three different things in both products: instructions you write, files you upload, and memory the assistant builds from your chats. Only the third changes by itself.
- ChatGPT has two modes, default memory and project-only memory. Claude gives each project its own memory and lets you switch off the account memory it would otherwise also read.
- Files and instructions are reference material for every chat in the project. Memory is a looser layer of notes, so put anything that must be exact in a file.
- Neither product shares a project with the other. Anything that has to follow you between ChatGPT and Claude needs a layer outside both.

In both ChatGPT and Claude, a project is a workspace with three separate parts: **instructions you write, files you upload, and memory the assistant builds from your chats**. Only the memory part changes on its own. The other two stay exactly as you left them, and every chat in the project sees them.

That difference is why people get surprised. They upload a brief and expect it to behave like memory, or they expect a project to remember a decision from last week and find it did not. This post explains how memory works inside ChatGPT Projects and Claude Projects, what carries between chats in a project and what does not, and how to set each one up. For the general mechanics, see how [ChatGPT's memory works](https://esment.notas.ai/learn/how-to-make-chatgpt-remember), how [Claude's memory and chat search](https://esment.notas.ai/learn/does-claude-remember-across-chats) work, and [AI assistants with memory](https://esment.notas.ai/learn/ai-with-memory) in general. Product details here are as of November 2026, and some are in flux.

## Instructions, files and memory are three different things

Before the products, the layers. They look alike in the sidebar and behave differently.

**Figure.** Instructions and files are reference material you maintain. Memory is notes the assistant keeps for itself.

| You control it | Assistant writes it |
|---|---|
| Project instructions | Project memory |
| Uploaded files | Picks what seems useful |
| Linked resources | Short notes, not transcripts |
| Identical in every chat | Can differ by chat |
Instructions tell the assistant how to behave: tone, format, what to assume. Files hold the material it should work from: a spec, a style guide, a transcript. Memory is the assistant's own shorthand about you and the work, written from conversations. If a detail has to be right every time, it belongs in a file or in the instructions, not in memory.

## ChatGPT Projects: default memory or project-only memory

In ChatGPT you can upload files, set custom instructions and link resources in a project, per [OpenAI Academy's guide to projects](https://academy.openai.com/public/resources/projects). Memory comes in two modes. OpenAI's [Help Center article on projects](https://help.openai.com/en/articles/10169521-projects-in-chatgpt) describes them as default memory and project-only memory, and we could not open that page directly, so what follows leans on Academy and on what other sources say it contains.

- **Default memory:** the project can use your saved memories from outside it.
- **Project-only memory:** your saved memories are not referenced in that project's chats. Chats can still reference other conversations *inside* the same project, never outside it.

Project-only memory depends on the personal switches: Academy says it requires "Reference saved memories" and "Reference chat history" to be on in your settings. If you want a client's project to stay clean of your personal preferences, this is the mode you want.

> [!NOTE]
> Several details are in dispute across sources, so we left them out as facts. Academy says project-only memory is for Enterprise and Business workspaces with Memory enabled. A third-party guide citing OpenAI's August 2026 release notes describes a general option in Project settings for unshared projects, and says shared projects are always project-only. Open your project's menu and see what your plan offers. Plan-by-plan file caps are reported but unconfirmed, so we do not quote them.

## Claude Projects: knowledge, instructions and per-project memory

A Claude project has [project knowledge and instructions](https://support.claude.com/en/articles/9517075-what-are-projects): you upload documents, text, code or other files, and Claude uses them as context for chats in that project. Instructions are set from the project page under "Set project instructions", per Anthropic's [guide to creating and managing projects](https://support.claude.com/en/articles/9519177-how-can-i-create-and-manage-projects). Projects exist on all plans including free, where you can create up to five.

Memory works as Anthropic describes in its [chat search and memory article](https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context): each project has its own memory. In a project, Claude uses that project's memory plus your memory from chats outside projects, but not memory from other projects. New memory formed in a project is saved to that project.

To keep a project fully separate, turn off **Use account memory** under Memory on the project's page. You can also turn off Memory for one chat from the "+" menu before the first message. Resetting memory from Settings permanently deletes all of it, project memory included, so use pause if you only want a break.

### What happens when the knowledge gets big

On paid plans, Claude switches to [retrieval for project knowledge](https://support.claude.com/en/articles/11473015-retrieval-augmented-generation-rag-for-projects) automatically when the files approach the context limit, which Anthropic says expands capacity by up to 10x. In practice that means Claude pulls the relevant passages instead of holding every file in view. Anthropic does not state the threshold, and we found no current official number for how much a project holds, so treat figures you see elsewhere as unconfirmed.

> [!WARNING]
> Anthropic says a new, beta version of projects is rolling out in stages to select Pro, Max and Ultra accounts, where each thread starts with the project's files, repositories, instructions and memory. If you are moved to it, existing users are emailed first. We could not verify how its memory behaves, so check what your own project shows.

## What carries between chats in a project, and what does not

Here is the practical version, with the caveat that memory is the loose part in both products.

| What | Carries into every new chat in the project? | Where it lives |
|---|---|---|
| Instructions | Yes, they apply to every chat | The project |
| Uploaded files | Yes, as context for the project's chats | The project |
| Project memory | Yes, when memory is on, as notes rather than transcripts | The project |
| Your account memory | ChatGPT: in default mode only. Claude: unless "Use account memory" is off | Your account |
| Another project's memory | No, in both products | That project |
| The other vendor's anything | No | Not shared |

Two limits are worth remembering. A new chat starts with a fresh window and sees the project's instructions and files plus whatever memory is on. In ChatGPT, chats can also reference other conversations inside the same project; we did not verify how far Claude's project memory goes beyond saved notes. And memory records what seemed worth keeping, so a decision made in passing may never get written down. For why that happens, see [why ChatGPT forgets](https://esment.notas.ai/blog/why-chatgpt-forgets) and [why Claude's memory seems not to work](https://esment.notas.ai/blog/claude-memory-not-working). If project memory is missing, run the [ChatGPT memory checklist](https://esment.notas.ai/blog/chatgpt-memory-not-working).

## Side by side: the memory layers in each product

| | ChatGPT Projects | Claude Projects |
|---|---|---|
| Instructions | Project instructions, set in the project | "Set project instructions" on the project page |
| Files | Uploaded or linked; caps unconfirmed | Project knowledge; retrieval on paid plans |
| Memory scope | Default or project-only | Project memory, plus account memory unless turned off |
| Sees your global memory? | Default yes, project-only no | Yes by default, no with "Use account memory" off |
| Sharing | Reportedly all users; shared projects project-only | Team and Enterprise, view or edit |
| Large knowledge | Not described in what we read | Automatic retrieval, paid plans |

## Set up a ChatGPT project that stays on topic

> [!STEPS]
> 1. Create the project and give it one workstream, such as "Lisbon app redesign", not a catch-all.
> 2. Write short instructions: audience, tone, format, and two or three standing assumptions.
> 3. Upload the stable material: the brief, the style guide, the glossary. Keep numbers and decisions that must be exact here.
> 4. Choose the memory mode. Pick project-only if the work should not pick up your personal preferences; keep default if it should.
> 5. Check that "Reference saved memories" and "Reference chat history" are on in your settings, since project-only memory depends on them.
> 6. After a few chats, read what ChatGPT saved and delete anything wrong.

## Set up a Claude project that stays on topic

> [!STEPS]
> 1. Create the project from **Projects** and give it one workstream and a clear name.
> 2. Open "Set project instructions", write the standing rules, and save.
> 3. Add project knowledge: the documents Claude should work from. Keep one file per topic, named plainly.
> 4. Decide on isolation. For a client or a sensitive topic, turn off "Use account memory" on the project page.
> 5. For a one-off chat that should leave no trace, turn Memory off from the "+" menu before the first message.
> 6. Review what the project remembers now and then, and fix it from the memory settings.

## When a project is not enough

Projects solve the problem of one workstream in one product. They do not solve the problem Alex Rivera, a product designer in Lisbon, runs into: the design decisions live in a ChatGPT project, the writing happens in a Claude project, and neither knows the other exists. Each vendor keeps its walls, and nothing documented crosses them. Moving context means copying files and instructions by hand, or keeping the facts somewhere both can reach. We cover that approach in [sharing memory between ChatGPT and Claude](https://esment.notas.ai/learn/share-memory-chatgpt-claude). If you are moving history rather than facts, [what transfers when you import ChatGPT chat history](https://esment.notas.ai/blog/import-chatgpt-chat-history) lists the limits.

> [!ESMENT]
> Esment is one memory that sits outside any single assistant: what you tell ChatGPT can be read by Claude, and the reverse, once both are connected. For work in a single assistant, projects plus built-in memory may be enough. See [connecting your assistants](https://esment.notas.ai/docs/assistants).

For the wider argument on why a fresh chat that knows nothing is costly, read [why AI assistants need memory](https://esment.notas.ai/blog/why-ai-assistants-need-memory). The short rule for projects: put what must be exact in files, what must be followed in instructions, and let memory carry the rest. That split is the core of [context engineering](https://esment.notas.ai/blog/context-engineering).

## Questions, answered

### Do ChatGPT Projects share memory with my main account?

It depends on the mode. With default memory, a project can use your saved memories. With project-only memory, it leaves them out and recalls only chats from inside the project, according to OpenAI Academy. OpenAI's pages disagree on which plans get project-only memory, so check the Help Center.

### Do Claude Projects see my other chats?

A Claude project uses its own memory plus your memory from chats outside any project, but not the memory of other projects. If you want a project fully separate, turn off "Use account memory" on the project's page.

### Is project knowledge the same as memory?

No. Project knowledge is the files and text you add, which Claude uses as context for chats in that project. Memory is what Claude writes down from your conversations. Both are scoped to the project, but you control the first one completely and the second one only loosely.

### What happens when Claude project knowledge gets too big?

On paid plans, Claude switches to retrieval automatically as knowledge nears its context limit, which Anthropic says expands capacity by up to 10x. You cannot trigger it by hand, and Anthropic does not publish the exact threshold. Free plans do not get it.

### How many files can a ChatGPT project hold?

We could not confirm it. Third-party guides citing OpenAI release notes report caps that rise with the plan, but those numbers disagree between sources and we could not open the Help Center page. Check your own account, and keep a project to the files it truly needs.

### Can a ChatGPT project and a Claude project share what they know?

No. Each vendor keeps memory and files inside its own product, and neither documents a way to hand a project to the other. To move context between them you copy files and instructions by hand, or use a memory layer that both assistants connect to.
