ChatGPT Memory: How It Works, What It Can't Do, and How to Build Your Own

Ask ChatGPT about a project you discussed last week and it may pick up where you left off, mention your preferred tone, or remember that you are vegetarian when it suggests dinner. That is ChatGPT memory, and since 2025 it has been one of the product's most-used and most-asked-about features. It is also widely misunderstood: people assume it is a database they can inspect, that it works everywhere, or that developers get the same thing through the API. None of those is quite true.

How ChatGPT Memory Works

ChatGPT memory is two mechanisms that OpenAI exposes under one setting.

Saved memories

Saved memories are discrete facts ChatGPT stores about you, either because you asked it to ("remember that I work in Berlin") or because it decided a detail was worth keeping. You can see the full list in Settings, delete individual entries, or clear everything. Saved memories are short natural-language statements, not structured records, and they are injected into the model's context on later conversations when relevant.

Chat history reference

Chat history reference lets ChatGPT draw on your previous conversations without saving an explicit memory. When it is on, the model can recall preferences, topics, and context from past chats and use them to shape new answers. You cannot see which parts of your history it is using; you can only turn the capability on or off.

Controls

• Memory can be switched off entirely, or saved memories and chat history reference can be toggled separately.
• Temporary Chat does not read from or write to memory.
• Memory is per account. Team and Enterprise workspaces have admin-level controls over whether it is available.
• Availability and limits differ by plan; free accounts get a smaller memory allowance than paid ones.

What ChatGPT Memory Cannot Do

For a personal assistant, the design is reasonable. Its limits show up as soon as you want memory to do more than personalize one person's chat.

• It is opaque. You can list saved memories, but you cannot see what chat history reference retrieved for a given answer or why.
• It is not exportable. Memory lives inside ChatGPT. You cannot move it to another assistant, back it up, or query it from your own tools.
• It is single-user and single-app. Memory belongs to your account. It does not model a team, a codebase, or a company, and it is not shared with other applications or agents.
• It is not in the API. When you call the OpenAI API, every request starts from zero. Developers building on GPT models have to supply memory themselves.
• It is not reasoned over. Saved memories are flat statements. If one contradicts another, nothing resolves it; the model simply sees both.

That last point matters more than it looks. Real memory changes over time: a preference is revised, a project is renamed, a decision is reversed. A list of statements has no notion of which one is current.

Why Developers Need Their Own Memory Layer

Because the API has no memory, every application built on an LLM faces the same question: how do you keep context across sessions? The common first answer is to stuff recent history into the prompt. It works until the history grows, at which point you are paying for tokens the model mostly ignores and still losing anything older than the window.

A memory layer solves this by moving memory out of the prompt and into a store the application controls. Information is written once, organized, and retrieved selectively when a new request needs it. For a deeper treatment of the options, see LLM memory and agent memory.

The strongest memory layers do three things ChatGPT's memory does not:

1. Model relationships. Facts are connected to the people, projects, documents, and decisions they concern, so retrieval returns a coherent slice of context rather than isolated statements.
2. Handle change. When new information contradicts old, the layer records which is current and keeps the history.
3. Serve many agents. One memory, readable by every application, agent, and tool in the system.

How to Add Memory to Your Own App

cognee is an open-source memory layer that turns text, documents, and conversations into a knowledge graph with a vector index on top. Adding memory to an application takes two calls:

import cognee
import asyncio

async def main():
    # Store: ingests, chunks, extracts entities, builds the graph
    await cognee.remember(
        "Maya prefers concise answers and is migrating the billing service to Postgres.",
        dataset_name="maya"
    )

    # Retrieve: routes the query to the best strategy and returns context
    results = await cognee.recall(
        query_text="What is Maya working on and how does she like answers?",
        datasets=["maya"]
    )
    for result in results:
        print(result.text)

asyncio.run(main())

Unlike ChatGPT's memory, everything here is yours: the graph is inspectable, exportable, and queryable; memory can be scoped per user, per team, or shared across agents; and contradictions are detected and resolved as new information arrives. The same memory is available to a chat app, a coding agent through the MCP server, and a background job, because it lives in a store rather than in one product's account.

If you are choosing between approaches, vector databases vs graph databases explains why cognee combines both rather than relying on similarity search alone.

How do I see what ChatGPT remembers about me?+

Open Settings, go to Personalization, and select Memory. Saved memories are listed there and can be deleted individually or cleared. What chat history reference uses is not shown.

Does ChatGPT memory work in the API?+

No. The OpenAI API is stateless; each request includes only what you send. Applications that need memory across sessions add their own memory layer.

Can I turn ChatGPT memory off?+

Yes. Memory can be disabled entirely in Settings, saved memories and chat history reference can be toggled separately, and Temporary Chat never uses memory.

What is the difference between ChatGPT memory and a memory layer like cognee?+

ChatGPT memory personalizes one user's conversations inside one product and is not exportable or available to developers. A memory layer is infrastructure your application owns: it stores memory as a knowledge graph, serves many users and agents, resolves contradictions, and is queryable from any tool.