The surprising similarities between AI and goldfish
How the smartest assistant can't remember yesterday
I've spent a large chunk of my career in Software engineering leadership roles. By far my favourite part has been mentoring and teaching. Helping others develop new abilities. Being there for the lightbulb moments when suddenly something clicks. I’m lucky to have had many chances to play a part in those moments.
Over these years I’ve lived the maxim "a little guidance goes a long way". Specific feedback to reinforce good habits. The occasional piece of critical feedback to correct mistakes. I do it all the time - it’s ingrained in me.
And I find myself treating AI similarly. I’ll give it a little advice to help it learn. And it seems to work. Tell Claude to be more concise and it complies. Give it a set of docs and ChatGPT is all too happy to emulate the style when writing new material. Ask it to pretend to be Steve Jobs and it'll answer like he would. It seems to learn.
But it's not the same. AI doesn't have long-term memory. At least not like humans. What it does have is amazing short-term memory combined with more knowledge than any human. But it doesn't have any domain specific long term knowledge. And it can't 'learn' - at least not in the conventional human sense.
To understand this better we need to dive into how memory works - both human and AI.
The nature of memory
The human brain forms memories through complex networks of neurons that strengthen their connections over time. When we learn something new, our brains physically change, creating and reinforcing neural pathways. This process - neuroplasticity - is why practicing a skill makes us better at it. The more we use a pathway, the stronger it becomes.
We have short-term and long-term memory. Short-term memory holds immediate thoughts and experiences - like remembering a phone number long enough to dial it. Long-term memory stores your experiences and knowledge - from riding a bike to your first day at school. New skills enter via short-term memory and, as we learn them, they get locked into long-term memory.
But AI memory works differently. Current AI models like Claude or ChatGPT don't actually form new memories during conversations. Instead, they have:
A training dataset - the "knowledge" encoded in the model weights during training.
A context window - their short-term memory during a conversation.
And zero ability to form persistent memories between chats.
When you tell Claude to be more concise, it's not learning a new behaviour. It's using its existing capabilities to follow your instruction within that single conversation. The next time you chat, it starts fresh. Like a goldfish’s memory, the short-term contents are gone.
And yet… …AI has the most amazing long-term memory. It knows far more than you or I could ever hope to learn. But this long-term memory is fixed. We can’t change it. Our goldfish is incredibly knowledgeable. It’s brilliant in the moment, but completely incapable of remembering yesterday.
Current attempts at memory
Our AI goldfish isn't completely without aids to memory. Some platforms are experimenting with limited forms of persistence:
ChatGPT supports basic long-term "memory", remembering selected facts about users. It's like having a cheat sheet. The model can reference some facts but can't build upon them - or develop new understanding.
Claude projects let you start chats with pre-defined knowledge, including preferences or facts about yourself. Our knowledgeable goldfish now has consulting notes from previous conversations.
These are useful features, but they're not true learning. Think of them as Post-it notes stuck to the goldfish's bowl.
What does this mean?
This forgetfulness brings a unique advantage. With humans, recovering from a difficult interaction requires careful navigation - apologies, explanations, rebuilding trust. With AI, you simply start a new chat. You have a reset button that returns everything to a happy place. It can be incredibly useful (imagine a world where we had something similar for humans).
And this doesn't make AI less valuable. Just different. You know exactly what you're getting each time. Each conversation starts from a clean slate, free from the baggage of previous interactions. There's a certain elegance to this simplicity.
Taken together there are profound implications for how we should work with AI:
First, don’t waste time trying to "teach" the AI.
Second, treat each interaction as fresh. Previous conversations don't influence the current one. Our goldfish starts each day anew.
Finally, use the reset capability to your advantage. Unlike human interactions, you can start fresh whenever needed.
AI can’t learn. But we can. So we need to adjust. The satisfaction of mentoring comes from watching permanent growth and development. With AI you're not teaching it. You're teaching yourself. Once you realise it is you who has to learn, you are on the road to harnessing AIs unique capabilities - and understanding its limitations.
As we continue developing AI systems, this distinction between human and artificial memory will likely remain crucial. Even as AI capabilities expand, the fundamental architecture of neural networks versus biological brains suggests we'll continue to see different patterns of learning and memory. Understanding these differences helps us build better tools and use them more effectively.
Like our brilliant but forgetful goldfish, AI's power lies not in what it can learn from us, but in what we can learn about working with a fundamentally different kind of intelligence. And perhaps that's the most valuable lesson of all.
Maybe. Even with today’s tech we could do a better job with memory. Store previous conversations in a RAG database, then search that and pull in relevant previous context for future chats. That’s likely how the current ChatGPT “memory” works - and probably what Mustafa Suleyman is referring to when he talks about “near infinite” memory coming next year.
It also raises some interesting questions:
Is near infinite memory universally better? Sometimes historical context is useful, sometimes it’s not. The human ability to forget inconsequential things may turn out to be a thing AI models need to replicate. Actually, I’ll go further - I reckon AI models _will_ need the ability to prune memory :).
Today models struggle to retain consistency within long chats - Gemini often forgets the original question, ChatGPT has a tendency to hit guard rails in long conversations and refuse to continue. If models struggle within the context of a single conversation, how will they cope with all the history as well? This one is probably soluable - Claude is very good at remaining consistent within a chat - maybe Anthropic have solved it.
Is memory the same as learning? In-context learning (provide lots of examples in the context) works. But I don’t want to have to train an AI model like I train my dog by repeating the same things over and over again.
Memory feels like it brings a new set of human-esque problems: things you want the model to always remember, things that are secret, things you don’t want the model ever to mention again - and the things that just don’t matter!
With increasing context window sizes, do you see the current as a temporary situation, where eventually the whole of the chat history will be context?