Just an Anonymous Room: Why ChatGPT is like Memento's Protagonist
Understanding AI Memory Through the Lens of Anterograde Amnesia
In discussions about large language models like ChatGPT, there's a lot of confusion around how they work - especially regarding memory and knowledge retention. Common questions include:
What does ChatGPT actually know about the world?
Does it remember our previous conversations?
Can it learn and retain new information over time?
Establishing the right mental model is crucial for understanding ChatGPT's capabilities and limitations. An insightful metaphor from a classic psychological thriller can help demystify how this AI remembers and forgets.
Christopher Nolan's neo-noir "Memento" (2000) centers around Leonard Shelby, a former insurance investigator searching for his wife's murderer. The twist? Lenny has anterograde amnesia, preventing him from forming new memories after a certain point in time. Like Lenny, ChatGPT operates with a peculiar memory architecture - it can draw upon a broad knowledge base, yet conversely suffers from severe memory limitations.
Let's explore how the uncanny memory condition of Memento's protagonist mirrors key aspects of how state-of-the-art language models like ChatGPT operate under the hood. (Minor spoilers ahead)
Early years: Foundational Models
Imagine Lenny growing up as an intelligent but naive schoolboy accumulating broad knowledge about the world from his 14 years - TV shows, biology, US government, etc. He can connect ideas, but lacks real-world feedback to apply that knowledge appropriately. For instance, behaving appropriately in a high-class restaurant would be difficult for Lenny.
The raw AI models behind ChatGPT start out like naive schoolboy Lenny - stuffed with a ton of book knowledge about how the world works, but no real experience applying it. They just learn the connections between words, which means they don't have any external guidance on how to produce good, safe or appropriate content. These raw models are called "Foundational Models", and you mostly don't see them in action unless you are working in AI development.
Lenny grows up: Fine-tuning
As a young adult, Lenny attends college and gets specialized training as an insurance adjuster. He receives feedback from teachers, friends, bosses and authorities, shaping how he thinks and acts. Besides the additional knowledge in business, he now also knows how to behave in tricky social situations.
After the initial training, Large Language Models go through "finishing school" - getting feedback from humans on what outputs are good or bad. This fine-tunes them to be more useful and well-behaved, like Lenny in his young adult years. With enough of it, the models become not only knowledgeable, but useful, polite, considerate and funny. They also learn to avoid dangerous topics such as bomb-building or hacking.
“So where are you? You’re in some motel room. There’s the key. It feels like maybe it’s just the first time you’ve been here, but perhaps you’ve been there for a week, three months. It’s kind of hard to say. I don’t know. It's just an anonymous room.” - Lenny
Amnesia sets in: The End of Model Training
In Memento, Leonard gets attacked in his home and after a hit on the head, loses his ability to make new memories. Everything he learns from that point on, he forgets. But he still keeps all his old knowledge, his memories and experiences before the attack. He is still Lenny, but frozen in time.
If you open up a blank chat session with ChatGPT, it will behave like Lenny: All memories from its foundational training and fine-tuning are there, but it has no new memories - regardless how much time has gone by or if you have interacted with it before.
Short Term Memory: Chat Sessions
Lenny is able to remember things for a short while - in the movie, it's about 5-20 minutes, depending on his stress level. Anything that happened before that, fades from his short term memory. This is not obvious to people around him: As a high-functioning amnesiac, he is able to muddle through, feigning memory through charm and intelligence.
“You fake it. If you think you’re supposed to recognize somebody you, you just pretend. You bluff it to get a pat on the head from the doctors. You bluff it to seem less like a freak.” - Lenny
Just like Lenny, if you are using a chat session to talk to ChatGPT for a while, it will remember what you've said up to a certain point. In the background of the session, anything you say to the Chatbot is fed back to the fine-tuned model along with a limited amount of previous conversation - about 3300 words for the free model 3.5 and 6500 words for the paid pro model 4. This means if you gave an information or instruction at the very beginning of a chat session, it will be gone after a text length of a long blog article or a short story. But even though ChatGPT doesn't have access to the memory anymore, it is convincing enough to act as if it was, guessing from the more recent context.
Starting a new chat session is like sleep for Lenny: He wakes up with all his memory reset, starting from zero. Everything from previous chat sessions will be forgot.
Polaroids & Tattoos: Custom Instructions
Lenny is aware that he will forget anything he learns. This is a problem, because new information is what he needs to solve the murder case. For this reason, he uses several crutches to keep his memory alive outside of his brain. He regularly takes Polaroid photographs of people, places and vehicles, noting a information like "Teddy 555-0134" or "She has also lost someone". He regularly checks the stack of photographs for reference, committing the information to his short term memory. Some really important facts he will tattoo on his body ("Never answer the phone").
Polaroids are for Lenny what Instructional Prompts are for ChatGPT Pro: With the paid version, you can use the "Customize ChatGPT" option to enter information into a "Custom Instructions" field. These instructions will always be at the forefront of the chatbot's memory. For example, you could instruct it to remember what languages you speak or in what style you prefer the answers. All of these instructions are semi-permanent, as you can easily delete them, just as Lenny can burn a polaroid or cover up a tattoo.
Maps & Phonebooks: PDFs and Data Lakes
But not every fact is important enough to get a tattoo or Polaroid: Some things Lenny just looks up. For example, he consults maps to drive around, phone books to call someone, or old police files to make research. Despite his memory condition, this is no problem for him, and he only needs to remember essential information from his lookup (i.e. a phone number, but not the whole phonebook).
Tools like ChatGPT can also look up information, using math to find the relevant facts. For example, you can upload a PDF into ChatGPT and query information from it. The chatbot won't remember everything that is in the PDF, it just will search it for the requested facts. Company LLMs can also be attached to a "data lake" containing proprietary documents. If you chat with such a tool, it will look up company-internal facts from these documents, just like Lenny looks up a fact in his police file.
The Big Picture
In summary, we have 5 ways Leonard and ChatGPT can learn or remember things. Here they are in overview.
Using this table, the Mememto metaphor can help us to answer the questions in the beginning:
What does Lenny / ChatGPT know about the world? Whatever they learned up to the accident / the end of its training.
Does Lenny / ChatGPT remember what has been said? For a short while, then they forget.
Does Lenny/ChatGPT know what's in my documents? They don't know, but if they have access to it, they can look up information inside the document. For ChatGPT, it needs the document uploaded or available, e.g. in a data lake.
Why does Lenny/ChatGPT forget the things I've told it to remember? Because they can't make new memories and will forget things you've said to them after a while.
Can you re-train Lenny/ChatGPT to learn new things? No, because they cannot make new memories. But you can give it memory helpers (Polaroids / custom instructions).
Conclusion
While the Memento metaphor provides an insightful framework for understanding ChatGPT's memory modes, it inevitably breaks down at certain points. Language models are highly complex mathematical systems processing information in radically different ways than the human brain.
For example, the metaphor doesn't cleanly capture how large language models are updated over time with newer versions superseding prior iterations. It's would be like Leonard getting an entirely new brain upload rather than building upon his existing memory architecture.
The metaphor also glosses over critical issues around potential biases, hallucinations, and the risk of outputting unsafe or untruthful information that language models can exhibit. Leonard may forget details, but he doesn't outright confabulate events that never occurred.
Despite these limitations, the core idea of ChatGPT having a broad foundational knowledge base but at the same time severe memory constraints is a powerful metaphor. When interacting with the model, it's wise to think of it as the world's most intelligent amnesiac.
So does ChatGPT actually "know" anything? Or is it simply an "anonymous room" bluffing convincingly based on its training? Use this framing as a guide, but don't fully commit to any certainties. To quote Leonard:
"Memory can change the shape of a room; it can change the color of a car. And memories can be distorted. They're just an interpretation, they're not a record, and they're irrelevant if you have the facts."
(This article was written in a symbiotic mode with ChatGPT 4 and Claude 3. The text is my own, the ideas were developed in tandem.)