Memory vs Project Instructions vs GPT Knowledge: ChatGPT's Three Context Mechanisms Compared

ChatGPT's three context mechanisms are three ways to make the model answer with background: Memory (cross-conversation personal memory), Project Instructions (project-level fixed instructions and files), and GPT Knowledge (a custom GPT's built-in knowledge base and instructions). They are often conflated, but their scope, maintainer, and lifecycle differ completely -- the symptom of choosing wrong is usually "it didn't remember what it should, and it dragged context where it shouldn't." This article compares the three on the same set of dimensions, then shows how to combine them.

1. One Table to Tell Them Apart

The three mechanisms answer three different questions: Memory answers "who are you and what do you prefer," Project Instructions answer "what must this project follow," and GPT Knowledge answers "what should this expert know."

2. Memory: The Cross-Conversation Personal Layer

The official help page's definition is direct: ChatGPT can pick up or remember details from today's conversation to reference in future conversations, making responses more relevant and helpful. Three key properties:

• Cross-conversation: today's chat influences tomorrow's answers, with no repeated background.
• Automatically accumulated: the model picks up details in conversation -- the essential difference from the other two mechanisms, which are configured by you.
• Manageable: memory follows the account and can be reviewed and managed in settings.

One easily missed read/write detail comes from the temporary-chat personalization mechanism (launched 2026-08-27): a personalized temp chat can read existing memories but will not create new ones -- which confirms from the side that Memory's read and write are separate things. If you want the model to "use your background but leave no trace," the default non-personalized temp chat is "no read, no write," and a personalized one is "read-only."

3. Project Instructions: The Fixed Background of a Workspace

Projects are ChatGPT's smart workspaces: group related chats, reference files, and custom instructions in one place. Two key mechanisms inside a project:

• Project-level Instructions: instructions you write once, applying to every conversation in the project -- "this project is a B2B redesign; here is the glossary."
• Reference files: files uploaded to the project, referenced by in-project conversations with the model citing relevant fragments and sources.

Memory within a project is shared across all its conversations and does not leak across projects -- which makes the project a "context isolation unit": your work project's background stays out of personal chats, and vice versa.

4. GPT Knowledge: A Distributable Knowledge Base

A custom GPT packages Instructions, knowledge files, and capability configuration into a publishable entity. The official documentation spells out the configuration surface: creating and configuring GPTs covers instructions, knowledge, capabilities, apps, actions, and version history; the knowledge-file limits are up to 20 files, each up to 512MB.

Its biggest difference from the other two mechanisms is distributability: once configured and published, teams or GPT Store users all get the same configured expert -- users do not need to (and cannot) change your Instructions and Knowledge. That fits "solidify a role's knowledge for many people," not "my personal, evolving preferences" -- that is Memory's territory.

5. How to Choose: Scenario to Mechanism

Combination is the norm, not the exception. A common team setup: a Custom GPT for the reusable expert capability, Projects for each long-running project, and Memory for personal preferences. Be deliberate about what each layer owns -- putting project standards into Memory, or hard-coding personal preferences into a GPT, are the classic mismatches.

Frequently Asked Questions

1. Do Memory and Project Instructions conflict?

Their scopes differ: Memory follows your account across conversations, while Project Instructions add a fixed background layer inside one project. In-project conversations get both -- project Instructions decide how to answer for this project, Memory decides who you are. To keep a project free of personal memory, use a default non-personalized temporary chat or override explicitly in the project.

2. Who maintains each mechanism?

Memory accumulates automatically by the model, which you can review and manage in settings; Project Instructions and project files are written and uploaded by you; a GPT's Instructions and Knowledge are configured by the creator -- once published, users do not need to (and cannot) change your configuration.

3. What are the limits on GPT knowledge files?

The official documentation is explicit: up to 20 files, each up to 512MB. Knowledge files are read-only background -- the model references them to answer but does not modify them. For a large, mixed knowledge base, prefer several focused GPTs over one universal one.

4. What should a team choose?

Long-running projects use Projects: chats, files, and Instructions live in one workspace with project-level memory shared inside. For a team-wide reusable expert role (like a brand copywriter assistant), build and publish a Custom GPT. The two combine: a shared GPT for the capability, Projects for per-project organization.

5. How do I keep sensitive topics out of memory?

Use a temporary chat. The default temp chat neither reads nor writes memory and stays out of history; as of 2026-08-27 you can enable personalization -- reading existing memories and custom instructions without creating new ones -- for "context continuity without traces." Saving a temporary chat converts it into a regular conversation.

6. Can the three mechanisms be combined?

Yes, and they should be. A typical setup: a Custom GPT provides the reusable expert capability, Projects organize each long-running project's chats and files, and Memory carries personal preferences across projects. Mind scope stacking: in-project conversations are affected by both Memory and project Instructions -- decide what each layer should own.

Next Steps

• Want to go deep on Memory? Read ChatGPT Memory: The Complete User Manual.
• Using Projects as a team workspace? Read ChatGPT Projects Complete Guide: File RAG, Custom Instructions, and Team Collaboration.
• Building a custom GPT? Read Building Custom GPTs: from idea to GPT Store launch.