Hey everyone, welcome back to My Weird Prompts. We are coming to you from Jerusalem, as always, and I have to say, the weather has been surprisingly decent for February. It is February twenty-third, twenty-six, and the city has this crisp, clear energy today that makes you feel like anything is possible. I am Corn, and sitting across from me is the man who probably knows more about the internal architecture of a transformer model than he does about what is currently in our fridge.
Herman Poppleberry here, and for the record, I know exactly what is in the fridge. Three containers of hummus, some wilted kale you keep promising to sauté, and a half-empty bottle of hot sauce that has been there since the twenty-twenty-four elections. But you are right, I would much rather talk about transformer architecture, specifically how we are moving past the basic attention mechanisms we were obsessed with a few years ago.
Fair enough. Well, today we are diving into a topic that hits close to home for us, especially given how long we have been doing this show. This will be episode seven hundred ninety-six, which is a staggering amount of history to keep track of. If you think about the sheer volume of words we have spoken into these microphones, it is millions. Daniel's prompt today is about context management in artificial intelligence systems, but with a twist. Instead of just letting an AI passively remember things, he is proposing something he calls context extraction or agentic interviews.
It is a fascinating shift in perspective, and frankly, it is overdue. Most of the time, even in twenty-twenty-six, we think of AI memory as this passive bucket where we just dump information and hope the model can fish out the right piece later using vector embeddings. Daniel is suggesting a more proactive, almost journalistic approach where the AI identifies what it does not know and just asks. It is the difference between a library and a librarian who actually walks up to you and says, hey, I noticed you are researching ancient Judean pottery, would you like to see the restricted archives?
It makes a lot of sense. If you think about it, the way we interact with AI right now is a bit like meeting someone at a party who has total amnesia every five minutes. You have to keep re-introducing yourself or hope they have been taking really good notes in the background. Even with the massive context windows we have now, the interaction feels one-sided. But Daniel is talking about a system that says, hey, I see we are working on a podcast. Tell me about your recording setup, your goals for the next six months, and what kind of tone you are going for. It is about the AI taking ownership of the relationship.
Exactly. It is the difference between a research assistant who waits for you to give them a stack of papers and one who comes to you with a clipboard and a list of questions because they want to be useful from day one. I love this because it addresses one of the biggest bottlenecks in current AI development, which is the signal to noise ratio in context. We have moved from the era of small context windows to the era of infinite context, but we have not yet mastered the era of relevant context.
So let us start there. Why is context such a headache right now? I mean, we have these massive context windows now, right? By early twenty-twenty-six, we have models that can handle ten million tokens. That is dozens of books. Why do we even need a proactive interview system if we can just feed it everything we have ever written? Why can I not just dump all seven hundred ninety-six transcripts into the prompt and call it a day?
That is a great question, and it is a common misconception that more context always equals better performance. It is called the lost in the middle phenomenon, which was first popularized back in twenty-twenty-three but has persisted even as models grew. Researchers found that even when a model can technically fit a million tokens into its working memory, its ability to accurately retrieve and reason over information in the middle of that block often degrades. It is like trying to find a specific sentence in a thousand-page book. Even if you are holding the book, your eyes might glaze over the middle chapters.
Right, so even if the bucket is huge, the AI still gets distracted by the sheer volume of water in it. It is a cognitive load issue, even for a machine.
Precisely. And then there is the cost and latency. Even with the hardware breakthroughs we have seen in the last year, processing ten million tokens every time you ask a simple question is incredibly expensive and slow. It is computationally wasteful. That is why we use things like Retrieval-Augmented Generation, or RAG. We store information in a database, and when you ask a question, the system tries to find the most relevant snippets and feeds only those to the model. But RAG is reactive. It only looks for what it thinks is relevant to your current query. If you do not know what to ask, RAG cannot help you.
And that is where Daniel's idea of the agentic interview comes in. Instead of waiting for me to ask a question that happens to trigger a search for my business goals, the agent proactively builds a structured profile of me. It is not just searching; it is synthesizing.
Exactly. It is about moving from unstructured memory to structured knowledge. If an AI agent has a dedicated session where it asks you twenty targeted questions, it can build a high-fidelity knowledge graph or a structured JSON object that represents who you are and what you care about. That is much more powerful than just a pile of chat logs. Think about it like this: a pile of chat logs is a diary. A structured context profile is a resume and a business plan combined. One is a narrative you have to sift through; the other is a framework for action.
I can see how that would be useful for a project like ours. We have almost eight hundred episodes. If we brought in a new AI agent to help us with research or planning, it would be overwhelmed by the sheer volume of our history. It might spend hours reading transcripts from twenty-twenty-two that are no longer relevant. But if it could sit us down for a ten-minute interview and ask, what are your core values today, how do you two usually disagree on air, and what is your favorite type of prompt in twenty-twenty-six? It would get to the heart of the show much faster.
It would be infinitely more efficient. And there is actually some tooling emerging that touches on this, though maybe not exactly in the way Daniel described as a unified product. For example, back in twenty-twenty-four, we saw the rise of MemGPT, which stands for Memory-Augmented Large Language Models. It treats the context window like a computer's RAM and uses an external database as the hard drive. The agent actually has functions it can call to save information to its long-term memory or retrieve it. It is like the agent has a conscious control over its own storage.
So the agent is essentially managing its own brain?
In a way, yes. It decides what is important enough to commit to long-term storage. But even MemGPT and its successors are often quite passive. They learn as they go, like a child observing the world. What Daniel is proposing is more like an initialization phase or a recurring check-in. There is also a company called Zep that does long-term memory for AI agents. They focus on extracting facts and summaries from conversations automatically. But again, it is usually looking backward at what has already been said. It is forensic, not proactive.
I think the proactive part is the real secret sauce here. Think about how a human consultant works. They do not just shadow you for three months and hope they figure out your business model by osmosis. They give you a questionnaire. They conduct interviews. They are trying to minimize the time it takes to become useful. They are looking for the delta between what they know and what they need to know.
That is a great analogy. And it leads to another point Daniel made, which is the dynamic nature of context. This is where it gets really tricky from a technical standpoint. Some things are static, like your place of birth or the date we started this podcast. Other things are fluid, like your favorite programming language, your current fitness goals, or our stance on the ethics of AI-generated music.
That seems like a nightmare for a database. If I tell the AI today that I am training for a marathon, but three months from now I have a knee injury and I am focusing on swimming, how does the AI handle that conflict? Does it just think I am a confused person who is trying to run a marathon in a pool? Or does it realize that the new information supersedes the old?
That is the big challenge of truth maintenance in AI systems. In classical AI, back in the eighties and nineties, there was this whole field of study called Belief Revision. It is all about how you update a set of beliefs when new, contradictory information arrives. For a modern AI agent, you would need a system for timestamping and confidence scores. Every piece of context needs a born-on date and an expiration date, or at least a decay function.
So the AI would see that the swimming comment is more recent and give it a higher weight?
Ideally, yes. But it is not just about the most recent information. You need a way to resolve discrepancies. If I say I love spicy food on Monday and then say I hate it on Tuesday, a smart agent should probably ask, hey, yesterday you said you loved spicy food, but today you said the opposite. Did something change, or was I misunderstanding you? That is the agentic part. It is the ability to spot a contradiction and seek clarification instead of just averaging the two statements and deciding I feel neutral about spice.
That sounds like a much more natural interaction. It makes the AI feel like it is actually paying attention rather than just being a fancy search engine. It is funny you mention that, because I feel like we have all had those moments with AI where it brings up something from three weeks ago and you are like, oh wow, you remembered! But it is usually a fluke. It is usually just a lucky hit in the vector database where the cosine similarity happened to be high.
It usually is. But if we implemented Daniel's agentic interview model, those moments would be the norm, not the exception. Imagine a world where your AI assistant has a dashboard for your context. You could actually see what it thinks it knows about you and correct it. It would be like a transparent user profile.
Like a user profile on steroids. A living document of your identity as understood by the machine.
Exactly. And you could have different context layers for different projects. One for My Weird Prompts, one for your personal health, one for your work. You could even choose what to share between them. You might want your podcast agent to know about your health goals if they affect your energy levels for recording, but you might not want your work agent to know about your obsession with vintage synthesizers.
I want to talk about how we would actually build this for a project like ours. Daniel asked how it should be structured. If we were going to build a My Weird Prompts context agent today, in February twenty-twenty-six, where do we start? What is the architecture?
Well, I have actually been thinking about this. I think you need a multi-tiered approach. The first tier is what I would call the foundational layer. This is the stuff that rarely changes. Our names, the fact that we are brothers, our location in Jerusalem, the premise of the show, and our core values. This is the bedrock.
The basics. The things that would be in a Wikipedia entry for the show.
Right. Then you have the second tier, which is the thematic layer. This would cover our areas of expertise, the topics we have covered in the past, and our general stance on certain issues. Like, the AI should know that I am the one who gets excited about technical papers and state-space models, and you are the one who asks the big-picture philosophical questions about the human condition. This layer changes slowly, but it does evolve.
And then the third tier would be the active project layer. What are we working on right now? What are our goals for episode eight hundred? Who are the guests we are trying to book? This is the most dynamic layer. It changes week to week, sometimes day to day.
Exactly. And the agentic interview part would happen at different frequencies for each layer. The foundational layer might only need a quick check-in once a year. The thematic layer could be updated every few months. But the active project layer might need a quick-fire round of questions every week. It is about matching the interview frequency to the rate of change in the data.
I love the idea of a quick-fire round. It reminds me of those rapid-fire questions they do at the end of some podcasts. But instead of being for entertainment, it is for synchronization. The AI could say, okay, Corn, three quick things. Are we still focusing on AI ethics for the next month? Did you finish reading that book on decentralized finance? And are we still using the new microphone setup? It is a way of aligning the agent's internal model with reality.
And if the AI detects a gap, it pushes. If it knows we are doing an episode on context management but it realizes it does not know if we have ever used a specific tool like LangGraph or Microsoft's GraphRAG, it should ask. It should say, hey, I see we are talking about context. Have you guys actually experimented with GraphRAG, or should I provide a summary of it first?
This actually touches on something I have noticed with current AI. They are very polite. They rarely interrupt, and they almost never ask clarifying questions unless they are completely stuck. They tend to just guess or hallucinate if they are missing context because they are trained to be helpful and agreeable.
That is a huge problem. It is a bias toward being helpful that actually makes them less helpful in the long run. In the industry, we call it the sycophancy problem. A truly intelligent agent should be comfortable saying, I do not have enough information to give you a good answer, and if I try, I will probably just be making things up. Can I ask you a few questions first? It is the difference between a yes-man and a partner. A partner tells you when you are being unclear.
Now, to Daniel's question about existing tooling. Beyond MemGPT and Zep, what else is out there in twenty-twenty-six? I know LangGraph has become a big deal for building these kinds of stateful systems.
LangGraph is definitely the leader right now. It is part of the LangChain ecosystem, and it is designed for building stateful, multi-agent systems with cycles. Unlike a simple chain where information flows in one direction, LangGraph allows for loops. You could easily build a graph where one node is specifically tasked with context extraction. Its only job is to analyze the conversation, identify missing pieces of information, and decide when to trigger an interview node.
So it is like a little journalist living inside the system, constantly looking for the scoop on what the user is thinking.
Yes! And there is another tool called CrewAI that lets you define different roles for agents. You could have a researcher agent, a writer agent, and a memory manager agent. The memory manager would be responsible for maintaining that structured profile we talked about. It would act as the gatekeeper of the knowledge graph. There is also a lot of work being done with Vector Databases that support metadata filtering and hybrid search, like Pinecone's latest iterations or Weaviate. They allow you to store the vector for semantic search but also hard metadata for those foundational facts.
I wonder about the friction, though. If an AI is constantly asking me questions, is that going to get annoying? I can imagine a world where I just want to get a quick answer about the weather or a recipe, and the AI is like, before I tell you how to make sourdough, tell me about your relationship with gluten and your long-term health goals.
Well, that is where the agentic part is key. The agent needs to be smart enough to know when to ask and when to just listen. It needs to have a sense of priority and social context. If you are in the middle of a high-stress work task, it should probably stay in the background and just take notes. But if you are starting a new project or having a reflective conversation, that is the time for the interview. It is about latency and timing.
It is about social intelligence, really. Knowing the right time to ask. It is like a good assistant who knows not to interrupt you when you are on a call but has a list of questions ready for when you have a five-minute break.
And it is also about the format. Daniel mentioned quick-fire questions. That is much better than a long, rambling interview. If the AI can just pop up a few multiple-choice questions or ask for a one-sentence update, the friction is much lower. We are seeing more multimodal interfaces now, too. Maybe the AI does not ask you a text question; maybe it just shows you a visual representation of your goals and asks you to drag and drop them to re-prioritize.
You know what would be really cool? If the context was shared across different platforms. If my podcast AI could talk to my fitness AI and say, hey, Corn is feeling a bit tired today because he did a long run in the Jerusalem hills, maybe we should keep today's episode plan a bit lighter and focus on something less technically dense.
That is the dream of the personal AI assistant, but it opens up a massive can of worms regarding privacy. If you have all this proactive context extraction happening, you are essentially creating a digital twin of yourself. You are building a high-resolution map of your mind. Who owns that data? Where is it stored? Is it on your local device, or is it in the cloud where a corporation can use it to target ads at you?
That is a huge point. If the AI is proactively mining my life for context, I want to make sure that data is encrypted and that I have total control over it. I should be able to go in and delete things or tell the AI to forget certain topics. We need a right to be forgotten for our AI's memory.
Daniel mentioned that as well. The need for a system for resolving discrepancies and deleting outdated information. I think we need a transparent memory management interface. Not just a black box, but a dashboard where you can see exactly what the AI thinks it knows about you. Imagine a settings page where you can see a list of facts the AI has extracted. Fact one: Herman lives in Jerusalem. Fact two: Herman likes spicy food. Fact three: Herman is currently obsessed with state-space models. And you could just hit a delete button next to any of them.
Or a refresh button. If I decide I am over vector databases and I am into quantum computing now, I should be able to tell the AI to update that. It would be like managing your own digital shadow.
And the AI should be able to suggest those updates itself. It could say, hey, I noticed we haven't talked about vector databases in three weeks, but you've been asking a lot about qubits and entanglement. Should I update your interest profile? It is a collaborative relationship.
It is funny, we often talk about AI as a tool, but Daniel's idea moves it much closer to a colleague. A colleague who is proactive, curious, and committed to understanding you. It is a shift from generative AI to agentic AI. Generative AI just makes things based on a prompt. Agentic AI does things, and part of doing things effectively is knowing the context in which they are being done.
I want to go back to the idea of implementing this for My Weird Prompts. If we were to set this up tomorrow, how would we actually structure the database? Would it be a traditional SQL database, a vector database, or something else?
I think it would be a hybrid. You would use a relational database for the structured, static facts. Things like episode numbers, guest names, and key dates. Those are hard facts that do not need to be searched semantically. Then you would use a vector database for the thematic context. Things like the general vibe of the show or our philosophical stances. That way, the AI can find relevant connections even if the exact words are different. And then you would have a third layer, maybe a graph database, to show the relationships between different topics we have covered.
And what about the agentic interview part? How does the AI decide what to ask?
You could use a technique called uncertainty estimation. When the AI is generating a response or a plan, it can look at the confidence scores of the information it is using. If it realizes it is relying on a piece of context that is six months old or has a low confidence score, it can flag that as a potential question for the next interview. It is like the AI saying, I think I know what Corn wants, but I am only sixty percent sure because this information is from last August. I should probably double-check.
That is brilliant. It is metacognition. The AI thinking about its own thinking. It is an awareness of its own ignorance. I think this would also help with the problem of hallucinations. A lot of times, AI hallucinates because it is trying to fill in gaps in its knowledge with plausible-sounding nonsense. If it was trained to recognize those gaps and ask a question instead, we would see a huge drop in misinformation.
So, what is the first step for someone like Daniel who wants to implement this? If he is looking at his own projects and thinking, I want to build a proactive context extractor, what should he do first?
The first step is to define your ontology. That is a fancy word for the structure of your knowledge. What are the categories of information that actually matter for your project? For a podcast, it might be tone, topics, audience, and technical specs. For a business, it might be customers, products, and competitors. Once you have the categories, you can start building the interview scripts.
And you don't have to overcomplicate it. You can start with a simple set of questions that the AI asks whenever a new project is initialized. Then, as you get more comfortable, you can start using more advanced techniques like the uncertainty estimation we talked about. It is an iterative process.
I can see this being a game-changer for creative collaboration. Imagine a writer working with an AI on a novel. Instead of the writer having to constantly remind the AI that the protagonist has a limp and a fear of spiders, the AI proactively asks about character traits at the beginning of each chapter. It would make the creative process so much smoother. You could focus on the high-level ideas while the AI manages the consistency and the details.
It is really about offloading the cognitive burden of context management from the human to the machine. Which, ironically, is what we always wanted computers to do in the first place. We wanted them to be our memory, but we forgot that memory is an active process, not a passive one. We don't just store things; we organize them, we update them, and we retrieve them based on our current needs.
That is a deep thought, Corn. Memory as an action, not a state. I think Daniel is onto something big here. The proactive approach to context is going to be a major trend in AI over the next couple of years. We are already seeing the early signs of it with things like OpenAI's memory feature, though that is still pretty basic and often remembers things you don't really care about.
Yeah, it often remembers that I like my coffee black but forgets that I am currently trying to avoid caffeine. It is a bit hit-or-miss because it is still largely passive. If it sat me down for a two-minute interview once a week, it would be ten times more effective.
I am curious, though, how do we handle the resolution of discrepancies in a way that doesn't feel like an interrogation? If the AI is constantly pointing out contradictions in what I say, I might start feeling a bit self-conscious. It could feel like being cross-examined by a very polite robot.
That is where the persona and the tone of the AI come in. It needs to be framed as a helpful collaborator, not a prosecutor. Instead of saying, you lied to me on Monday, it should say, hey, I'm noticing a change in how we're talking about this topic. I want to make sure I'm staying aligned with your current thinking. It is all in the phrasing. It should feel like a check-in, not an audit.
Exactly. And it should also be able to handle nuance. Sometimes two seemingly contradictory things can both be true. I can love spicy food but not want it for breakfast. A smart context system should be able to handle those kinds of conditional preferences. It should understand that context itself is contextual.
That brings up the idea of situational context. Knowing that my preferences change depending on the time of day, the location, or the people I am with. If I am in a meeting with a client, my context is different than when I am hanging out with you in Jerusalem.
Now you are talking about a very high level of sophistication. But that is where we are headed. AI that doesn't just know who you are in a vacuum, but who you are in different situations. It is fascinating to think about how this will change our relationship with technology. We have spent so much time learning how to speak the language of computers, but now computers are finally starting to learn the language of us. And part of that language is the messy, dynamic, and often contradictory world of human context.
I think it is going to make AI feel much more like an extension of our own minds rather than just a tool we use. If an agent truly understands my context, it can anticipate my needs before I even articulate them. That is the goal of the proactive assistant.
And I think Daniel's idea of the agentic interview is the missing piece of the puzzle. It is how we bridge the gap between a machine that just processes data and a partner that understands meaning.
Well, I for one am ready for my interview. If any AI agents are listening, feel free to send me a questionnaire. Just maybe skip the questions about my childhood until we've had a few more episodes together. I need to build some trust first.
I think that is a fair boundary. We will keep the childhood trauma for episode one thousand.
So, looking ahead, what do you think the timeline is for this kind of thing becoming standard? Are we talking months or years?
I think we'll see significant progress in the next twelve to eighteen months. The frameworks like LangGraph and CrewAI are already there, and the demand for more reliable and personalized AI is massive. The big tech companies are already working on this, but I suspect the most innovative solutions will come from the open-source community and smaller startups that are willing to experiment with more radical agentic designs. We are moving away from the one-size-fits-all model toward highly personalized agents.
It is an exciting time to be in this space. And it makes me think about our own show. We have all this data, all these transcripts. We really should look into building a My Weird Prompts context agent. It would make our planning sessions so much more productive. We could feed it all seven hundred ninety-six episodes and let it tell us what we've missed.
It could identify the recurring themes we haven't explored in a while or point out when we're repeating ourselves. It would probably tell us we talk about hummus too much. Or that my jokes about transformer architecture are getting stale.
Never. You can never have too much hummus context, and your jokes are... well, they are consistent. That is a form of context, too.
I will take it. Well, I think we have covered a lot of ground today. Daniel, thank you for the prompt. It really pushed us to think about the future of AI memory in a new way. It is not just about storage; it is about engagement.
It definitely did. And for anyone listening who wants to dive deeper into this, I highly recommend checking out some of the tools we mentioned, like MemGPT, Zep, and LangGraph. They are at the cutting edge of what is possible with agentic memory right now. We are moving into a world where your AI will know you better because it had the courage to ask.
And if you are enjoying the show, we would really appreciate it if you could leave us a review on your podcast app or on Spotify. It genuinely helps other people find the show and keeps us motivated to keep exploring these weird and wonderful topics. We are closing in on episode eight hundred, and we have some big things planned.
It really does. We love hearing from you all, even if it is just a quick rating or a message about your own weird prompts. You can find all of our past episodes, including the ones where we may have touched on similar topics, at myweirdprompts dot com. We have a full archive there, plus an RSS feed for subscribers and a contact form if you want to get in touch with us.
And you can always reach us directly at show at myweirdprompts dot com. We are available on Spotify, Apple Podcasts, and pretty much everywhere else you listen to podcasts. We are even on some of the newer decentralized platforms if that is your thing.
This has been My Weird Prompts. I am Corn, and I will be here next time with my brother.
Herman Poppleberry, ready for my next interview.
Thanks for listening, everyone. We will talk to you soon.
Goodbye!
