Latest / TeqTalk with Jas / The Smartest AI on Earth Forgets You Every Morning?
Transcript
- Teqfocus: Meet Nora. She is a chief operating officer. Runs operations for a company with 3,000 people. And last Monday she hired the smartest assistant in the world. It reads faster than any human. It never sleeps. It can draft a board memo in nine seconds. There's just one problem. Every single morning the assistant forgets she exists. Monday, Nora tells this assistant everything. Her priorities, her team, the deal that is falling apart. Tuesday, when Nora says, Good morning, it replies, Hello, and who are you? This isn't a bug. This is every AI today. The smartest machines we have ever built have the one flaw that makes them almost unusable. They cannot remember. So this video answers. One question. And by the end, you will understand it better than most people in the industry. Why can't AI remember you? And what changes the moment it can? I am Jazz and this is Tech Talk. Tech Talk isn't just a podcast, it is an executive conversation series where the leaders actually shipping enterprise AI break down what's real and what's hype. Paired with the deep dive episodes like this one. Where we unpack the concepts underneath. And this episode has a backstory. In my last conversation, a leader and I went deep on AI memory and the questions afterward wouldn't stop. Let's go back to Nora's assistant. We will call it Eva. Eva is brilliant, but Eva has amnesia. And to understand why, you have to understand one thing about how these AI Word. Everything Eva knows in a conversation lives in something called as context window. Think of it as a whiteboard. Eva can read everything on that whiteboard instantly. She is a genius level with whatever is written on there. But the whiteboard has an edge, and when it fills up, the older stuff gets wiped to make some room. Conversation ends, the whole board gets erased. That's The amnesia. Ava doesn't have a bad memory. She has no long-term memory at all. Just a whiteboard that resets. And here is the part that surprises people. You might think just give Eva a bigger whiteboard. But a study called Lost in the Middle found something white. When you stuff important information into the middle of a long context, accuracy drops by more than 30%. The model basically skims past it. So the bigger whiteboard isn't memory, it's just a bigger surface to lose things on. Here is the first principle the whole field rests on. It comes from the koala paper out of Princeton. Language models are stateless. You prompt the model, it answers and it forgets. The next call starts from zero. The model never remembers anything ever. So here's the twist. The memory was never in the model. It is an architecture we built around the model that forgets. And this is a wall the entire industry hit. Because you cannot build an agent, an AI that takes real action over days and weeks and on top of something that forgets. Autonomy without memory is impossible. Because the people solving this didn't invent something new, they looked at the best memory system that already exists, which is Yours. Your brain doesn't have one memory. It has several and they do different jobs. And to make Eva useful, Ingenious had to give her the same ones. First, the working memory. That's the whiteboard. When you are holding in your head right now, as you listen to me, Eva already has this. It's just not enough. Then come into the long-term memories. There are three kinds of it. One, episodic memory, memory of events. You remember what you had in the breakfast. Eva needs to remember that yesterday Nora said that Henderson deal is the top priority. Specific moments tied to time. The second is semantic memory. Memory of facts. You know, Paris is the capital of France. Without learning it, Eva needs to know Nora's company has 3,000 employees that her CFO is named Marcus. That the fiscal year ends in March. Just Knowledge detached from any moment. The next one is procedural memory. Memory of how. You don't think about riding a bike, you just write. Eva needs to learn how Nora likes things done. Her tone in emails, how she structures a memo, and to do it automatically without being told each time. That framework, working, episodic, semantic, procedural, is exactly what Lays out. AI memory isn't one thing, it's a system. And every piece we add makes Ewa a little less of a stranger every morning. Now the question is how the memory actually works. This is the part that feels like real magic unless you see it. And then it is beautiful. When Nora tells Eva something important, Eva doesn't save it like a document in a folder. Folders Need labels and you also have to remember that labels to find it again. Too slow and too brittle. Instead, Eva converts the meaning of what Nora said into a kind of coordinate. A location in a vast space where ideas that mean similar things sit close together. The Henderson deal is at risk. Lanceer, we might lose Henderson, even though they don't share the same words. This space is called as a vector database and the coordinates are called embeddings. And why does it matter? Because now when Nora asks weeks later, what were we worried about that client? Eva doesn't search for the matching words. She searches for matching meaning. She travels to the neighborhood of that space and pulls back everything nearby. This whole move. Store by meaning, retrieve by meaning has a name and it is retrieval. It's the RN rag if you have heard of the term. It's how Eva reaches a past whiteboard into everything she has ever stored, grabs only what is relevant, and puts it on the whiteboard right when it is needed. And the best systems add one more layer. They do not just store everything that's noise. Eva reflects. After a conversation, she decides what actually mattered, summarizes it, and keeps that. Just like you do not remember every word of a meeting, you remember what counted. Write, retrieve, reflect. That loop is the engine. And the moment Eva has it, everything changes. Why this unlocks everything? Because now, watch what Nora's assistant becomes. Week one. Eva needed to be told everything twice. Week four, Nora opens her laptop and Eva says, The Henderson contract is up for renewal in 11 days. Last quarter, you flagged their payment delays. I have drafted three options. Want to see them? Nobody asked her to do that. That's the leap. Memory doesn't make Eva convenient, it makes her autonomous. Because an agent that remembers what it did yesterday. Can plan for tomorrow. It can catch its own mistakes. Last time I sent this at 5 p.m. and got no reply. Let me try this morning. It can build on its own work instead of staring from zero every single time. This is the difference between a tool and a teammate. A tool waits for instructions. A teammate remembers the goal, remembers the history, and moves things forward on its own. Nora didn't get a smarter chat bot, she got a chief of staff. One that started as a stranger every morning and became the person who knows the business better than anyone in the room. That is what agentic AI actually means. Not a bot that answers, but an agent that remembers, learns, and acts. And memory is the thing that unlocks all of it. Now this sounds very interesting. Let's talk about some of the risks. Think about what Ava knows now. Everything, every deal, every worry Nora ever voiced, every weakness in the org. That's incredibly powerful and incredibly sensitive. Whoever controls that memory controls a mirror of the entire business. Privacy stops being a checkbox and becomes the whole game. Then there is a danger. What if Ava remembers something wrong? Say she misheard a number early on and stored it. Now she doesn't just repeat the mistake, she builds on it. So bad memories compound. A wrong fact stored confidently can quietly poison every decision downstream. And because Eva sounds so sure, nobody catches it. And finally, the one that runs the whole show, which is cost. Storing memory, searching it, pulling it back thousands of times a day, none of it is free. There's a real reason your AI doesn't remember everything yet. Sometimes it's not that they cannot, it's that remembering at scale is expensive. So the frontier isn't just give AI more memory, it's the right memory. Accurate, private, affordable, and that's the problem. It's very much unsolved. It's exactly what the sharpest teams in the world are racing on right now. So let's go back to Nora one last time. She hired an assistant that forgets everything in the morning and she ended up with a chief of staff that knew her business. And here is the answer to the question we opened with. Nothing about the intelligence changed between Monday and week four. The only thing changed was the memory. That's the whole story of where AI is going right now. We spent years making these models smarter. The next race, the one that actually matters, is making them remember. And this is exactly why I sat down with the leaders building it. Concepts like these are in theory. They are being shipped inside real enterprises today with real stakes. If you want the executive view on AI memory, go watch my conversation with the leader who kicked all of this off and it's linked right here. This educational deep dive and the conversation are the two halves of the same story. That is what TikTok is: the concept and the people putting it into practice. One last thing before you go, I want your honest answer in the comments. Would you actually want an AI that remembers everything about you? The good, the bad, every mistake, every conversation? Is the best assistant you have ever had? Or does it genuinely scare you a little? Tell me below, I read them and the best answer might make it into a future episode. If this made AI memory, finally click subscribe. Because every episode here is exactly this. A real executive conversation plus the concepts underneath it. That's Tech Talk. I am Jess. See you all in the next one.