Build an AI Content Research Agent with Claude Code & Oracle AI Database

Hello everyone and welcome to today's session. My name is Rhys and I'll be your moderator for today. We are going to get started at the top of the hour in just over 3 minutes or just under 3 minutes, apologies. So then if you would like to join in in today's session, we do have a little bit of setup info that is required. You will have noticed this in the reminder email that you got sent about an hour ago. So go have a check of that. However, the setup info is in the session resources document that we've got pinned in the chat and in the video description. If you don't see it in the video description, please reload the page and it'll appear there. Yeah, check the pinned link in the chat to find the link to the resources document. And then the first link in there, we've got a GitHub repo which has got everything you need. Obviously today's session, we are going to be using Claude Code, so you will need access to Claude Code if you do want to join in with us. And yeah, if you'd like to set up yourself with an Oracle database as well, there is the instructions to do so in the GitHub repo in the session resources. So please do check that out. Check the pinned link in the chat if you'd like to check that out. However, if you would just like to watch along today, please do register for the session using the QR code that's on screen now. That means we can email you tomorrow with the recording and everything else that you need to complete this in your own time as well. If you have any questions at any point throughout the session, let us know in the chat. We'll be answering your questions for the last 10 minutes of the session, so make sure that you stick around for that.

Brilliant.

I'll repeat these messages for any new joiners very shortly, but yeah, please do check out the session resources, the pinned link in the chat, and register for the session if you haven't done so already. And for anyone that's just joined us, well, yeah, thank you for joining us. We're going to get started very shortly and yeah, we'll be off pretty soon. There is a little bit of setup for required for today's session. If you would like to join in live, the info for that is in the session resources document. That is the pinned link in the chat. So go check that out if you'd like to join in live. However, we are going to be sending the recording to everyone that's registered for this session. So please do scan the QR code that's on screen now or head over to datacamp.com/webinars. Where you'll find this session as well as all of our future sessions as well. In terms of the tools we're going to use today, obviously we're going to be using Oracle, uh, for our database, and we are going to be using Cloud Code. So if you don't have access to those, uh, get those set up. There is info on, uh, the Oracle bits in the session resources. That's the pinned link in the chat at the moment. Also, if you'd like to catch up with yesterday's session, there is a link to that in the resources document as as well. We've also got the slides that we're going to be covering today in there too. So yeah, everything you need is in the resources document that's pinned in the chat. And if you have any issues, let me know in the chat. I'll be keeping a keen eye on it. If you have any questions throughout the session today, let us know in the chat. We're going to be answering your questions for the last 10 minutes of the session. Make sure you stick around for that.

Brilliant.

I think we are ready to get going. So now I'll hand you over to your host for today's session. Richie, Richie, please take it away.

Hi there, Data Scamps and Data Champs. This is Richie. Welcome back to the Anthropic Age. So Claude is the hottest large language model right now. We've got a whole week of sessions to help you make use of it. So yesterday we looked at Claude Cowork. This is a no-code solution for automation. Today we're making use of Claude Code. This is aimed at software developers and other people who are more amenable to using code. We're gonna be creating an agent for research. Since this agent needs to remember things, we'll also be making use of Oracle's AI database. So Claude plus Oracle, great combination. Our guest is Cassius Sabanda-Lee. He is an AI developer advocate at Oracle. So welcome Cassius, great to have you here.

Thanks for having me, Richie Cotton.

Yeah, I'm very much looking forward to this session. So Cassius is a data engineer turned AI engineer turned educator. So he trains people on how to use and build with AI. Before his time at Oracle, Cassius was a generative AI engineer at Accenture, where he built an award-winning AI tool for HMRC. That sounds pretty cool. All right, wonderful. I'll let you take it away, Cassius.

Cool.

Thank you very much.

So hi everyone. My name's Cassius and I'm an AI developer advocate at Oracle. And today I'm going to be running you through a limitless workflow that I've been working on for the past few months. Uh, my role as a developer advocate basically sees me trying to understand loads of different concepts in AI. And as you probably know, there's so many things happening every single day. Um, and it's pretty hard to keep up. So I found that I really struggled with just reading a lot of different information from so many different sources. So I thought about making something that I would use myself to help me build mastery in these different, these different areas. So today I'm gonna walk through my memory-powered learning tool and it's built on Claude Code. And it's built using Clore Code. And Oracle AI Database is the memory core. The whole basis behind it is that you can turn pretty much any concept into a mastery of that subject. So, and then after you've gained that mastery, then you can turn that into content. So it's going to be about a 45-minute session, 2 live demos, and there's some QA at the end. So if you do have any questions, you can just, I guess, put them in the chat and Yeah, we'll go from there. So, and to say as well, if you have set up the, I think a document was sent out earlier how to actually set up and get the GitHub repo. You can do that. Don't worry if you haven't because I'll kind of go through everything here and I will be giving the entire, I suppose, workflow just free at the end. So you can have a look in your own time if you just wanna watch. So let's start with what's in it for you today. We are building a system that basically allows you to actually retain what you study because we know there's a lot of different concepts and topics out at the moment, especially within AI. There's a new framework or technology every few weeks, so, or every few days actually. So we really want to be able to retain what you study. So we want a system that forces real understanding, not just generated outputs. And then on top of that, I want— we want to create a system that actually understands what you understand and, and can actually see what you don't as well. Um, and that will show that later on with a knowledge graph. And then the third section is turning that mastery of a concept into content that you can ship. So that could be, uh, you know, a blog or, uh, some sort of script, anything you can really think of. And just on the left here, you can see that later on I'll show you, we've— I've actually added an Obsidian command center front end that we can actually show, and you can see the concepts you understand, you don't, and we'll go into that later on. So I think it's best to pull back a bit and start at the ground floor, and we'll talk about actually what is Claude Code. So if you don't know about Claude Code, All it is is literally a reasoning model that's kind of— that understands the breadth of human knowledge and it's running inside your terminal. So you can type any sort of any language but into any language into a prompt. And what Claude Code will do, it will reason and figure out if it needs to pick tools, call them, read the result, and then decide what to do next. And that's the basis of the agent loop. So if you see the diagram at the bottom, You can see that number one, you'll type something, you'll type a prompt in your own natural language. Then it'll go around and Claude Code will think about your prompt and basically turn that into, turn that reasoning into, does it call a tool, which could be reading files on your computer, running a bash command, actually just searching the internet or calling a database. And then once it's actually called that tool, it will read that result and figure out, do I need to continue or should I just stop here and give the user an answer? And that's the basis behind pretty much every, um, agent you can think of. Um, and what I've, if you want a deeper understanding, this is just a quick one, but if you want a deeper understanding, there's a deep dive, uh, blog that I've created on the Oracle Developer Blog. And it's basically about what is the agent loop. So you can look, um, a bit more there. There's also the QR code. So you can go into that. And also, just to reiterate, this is the— Claude Code is the runtime that this Limitless workflow works under, which we'll see later on. And then another really, I suppose, interesting new thing in Claude Code recently, or the industry, is a skill. And you might have seen this on social media recently. There's all sorts of skills going around that can help you do specific tasks a lot more effectively. The basis of a skill, all it is, is a folder with a markdown file in it, a skill markdown file, and it contains a name, a description, and a few instructions. And what Claude does when you type in, you know, run this skill, Claude reads that when it's relevant and follows it like a playbook. So on the left-hand side here, I've got an— I've got a high-level example of what a skill would look like. So this is, uh, one of the skills we'll be using today is the XP Builder, which we'll go into, uh, and you can see it's got the name, the description, and then some basically, I suppose, tasks that it would actually do and rules and things. So it's a skill is a procedural workflow that Claude Code uses to do a task. So I'll go into the next slide. And just to reiterate, this is one of the skills that we'll be using today in action. So I thought it would be good to show off. An example. And where did I get the idea for this Limitless workflow? So one of my favorite films, uh, as someone with diagnosed ADHD, this really captured my attention. It's, um, it's basically called Limitless, and the basic premise behind it is this guy Eddie Morra, he takes this experimental new drug called NZT-48, and his brain wakes up. He can learn new languages in a weekend, he has total recall of all of his memories, and he can basically plan 10 steps ahead of anyone. I mean, this is a film, so we don't actually have that, uh, that pill, but we do have these really powerful reasoning models and a durable database and the tools that we can build a sort of replacement in a way. So in short, we don't have NCT, but we do have AI. Um, so that's kind of where, uh, I'm basing this, this idea on. And I wanted to go into the core issue I see with learning with AI at the moment. I feel like, and this is my own theory from what I've read, is that we know that AI makes output really easy and you can do so many new things. Your productivity is through the roof. So many new things that you wouldn't be able to do on your own. But I feel like it makes— it doesn't make understanding automatic. And the issue kind of with this is I believe because AI is— we've got such access to AI that we can— any sort of, uh, problem we've got, we can ask AI. But our brains don't go through that mental process of like actually struggling to find out the, the actual answer or learning the answer. It's just like given to us. So I feel like that's the issue there, is we're not actually building, um, that encoding into our brain. And this is actually backed by evidence. So, um, I'll explain one study by Bestani et al., which they basically found that unrestricted AI builds an illusion of learning. So in this study, they had students use AI to basically, uh, learn different subjects that they were supposed to be, uh, in, in an exam. And they found that while they were using AI in practice, the students saw a 48% increase in like the problems they were able to solve. And they did really feel like they were learning, which is my constant day-to-day. I feel like I'm learning so much. But when the students were actually in the real exam without the AI, those same students, um, saw lower— 17% lower scores than their peers who had never even touched AI. So that kind of shows that you— there is that illusion of learning in one way. And are we actually building those mental pathways? I don't know, we'll see. Also, I've actually, uh, you can check the QR code there because the study is right there if you check. So the learning science has been ambiguous for decades, and there's mainly 3 levers that do most of the work. Number 1 is retrieval practice, and the basis behind this is retrieving, uh, information from your memories is a much stronger encoder than rereading something or watching something. So if I was to say, um, ask you a question, and the thought of you actually having to bring that out of your memory is a lot stronger than just rereading something or cramming in information before an exam. It's best to have that back and forth conversation. We've also got self-explanation. So you being able to explain a concept in your own language forces deep encoding into your brain, and it actually helps you surface the gaps that you, that you don't understand. If you're having a conversation with someone and you're having to explain something, then you can kind of understand what you know and what you don't know. And then we've got the Protégé Effect. I might be saying that wrong, but, uh, and this is basically teaching, teaching anything to someone. So we actually can consolidate knowledge better when we know we have to explain it to someone else. So for an example, if I was to say read a 10-page research document right now, you might read it and you might think that you're taking in all the information, um, but you probably won't. But if I was to say, uh, you know, you've got this 10-page document, in 10 minutes you've got to explain that to somebody else, then that's when your brain starts to consolidate information a lot more, um, effectively, we'd say. We're nearly at the end of the PowerPoint before the demo. So I'm definitely powered on here. But I want to explain a bit about how human memory maps to agent memory. So, hmm, I'll first explain the different parts of the human memory. So human memory isn't one thing. And there's three fundamental types and Oracle Database can map to all these three. So firstly, we've got episodic memory, and this is just what happened. So you can think of it as like a specific event that happened in the past. So, uh, you know, yesterday's session with, uh, with an AI agent, or the, the meeting you had last Thursday or something like that. And the way the Oracle database interacts with this is it saves every session of these learning sessions that we're having with AI, um, what happened to the session, what topics we covered and then has like a diary in the system that it can reread later on. So it actually has that memory in there. You've then got type 2, which is semantic memory. And that is kind of what is true. I say true in quotes. That's facts and meaning. So, and that's decoupled from when you actually learnt them, but it's the kind of the shape of your knowledge. So For an example, it's like thinking that, you know, the sky's blue. We all know that it's kind of like a fact. That would be semantic memory. And within the Oracle database, we have kind of like a report card that updates itself of like what it sees as true as we build out this understanding. And then to finish off, we've got procedural. So that's how you do things. And when I mentioned before about Claude skills, That is kind of procedural memory. It's how you do things, the different workflows and tasks that you do, and the patterns that you kind of do. It's kind of like you'd think of muscle memory. So the procedure of riding a bike, you pedal the wheels and I suppose move the, move the handlebars or beep the horn, whatever. And that's the same thing with— sorry, in Oracle, I had AI database, we can capture that procedural memory as well. Let me just go on to the next. And just as a reminder, we've also, um, created another deep dive on what is agent memory, and that's in the Oracle Developers blog, which you can see right there at the QR code. Last one. So, uh, the solution for creating a learning tool As we saw in the, uh, in the evidence, you can't just give, uh, people the answer because then they don't have the, the mental, uh, kind of hardship of actually remembering that stuff. So principle 1 of this is that we use hints and not answers. So we use a Socratic tutor that basically probes you, reframes things, escalates, um, and kind of like makes it so you're doing the cognitive work, not just it's auto-generating something and you're just reading it. And then there's also the fact that we want this to be spaced and repeated. So one session with this, with this, um, kind of tool will not build a mastery in whatever. To actually learn something, you have to repetitively go over that thing and, and, you know, uh, make those you have to build the understanding in, in a way. Um, so retention compounds over time. So we basically need to do this over days or weeks, and that's how you actually build a really good understanding of a certain topic. So now we're going to go into the first demo, and I'm going to show you how to basically, uh, run this XP Builder and see what we can actually do. So I want you to look for how it teaches and how it doesn't just like give you the answer even when I kind of push it. It's a— this is a real back and forth conversation that you would have in a certain time period. So if you just bear with me one second, I'm just going to swap over the screens. Right. First try and I got it right. Perfect. So Uh, this is VS Code. If you don't know, it's kind of like a— it's an IDE, so you can, uh, basically understand your code base and you can run different, different tools. And the main tool that we're going to be using today is the terminal. So, uh, let me just— actually, one second. Um, just to— just so you understand, I'm actually just running this within my Limitless kind of, um directory. So once you're in the certain directory that you want to, you want to use, you can literally, to open Claude Code, all you have to do is type in Claude. Oh, Claude. There we go. And then this is your kind of interface for Claude Code. So you can see that we've actually quite good timing this, this webinar because Opus 4.7 has just released. So I've been testing that over the past few, past few days. And it's, it's pretty good. Um, so to run this XP Builder, all you have to do is type run XP— oh, on agent memory. So I'm just going to write run XP Builder on agent memory, and we will see what happens. So As I mentioned before, all you have to do to actually— once you've got a skill installed, all you have to do is reference that in the prompt. So because I've put /x run xp_builder, Claude can understand that I want it to run the XP Builder from this directory, which is a Python script that we've, that we've put in there. Um, so it shows— let me just roll back up— it shows here that it's successfully loaded the skill, and then we're kind of now in the session. And it's also going to give me some other information about how, um, you know, the session key, which is basically the transcript that we have in this session right now, um, and, uh, the state path, which is basically where this is going to be kept, the JSON, uh, JSON file where it's going to be kept before it goes into the, the database. Oh, sorry, one second. Um, so to explain, we've got a, um, we've also got a heads-up display in the Limitless Builder. So I'm going to go through these line by line so you can kind of understand, uh, so when we actually have this, when we have this conversation back and forth. Number 1, this basically explains that we're using the Limitless XP Builder. Um, the topic we're trying to understand right now is agent memory, and we're actually in the diagnostic phase. So at the start of each chat it'll be in this diagnostic phase where it's asking you questions back and forth to kind of understand what you know and what you don't know. And as you build through this, uh, as you continue to use this and have different topics that you put into the database, it's going to have like a kind of map of your own brain of what you know and what you don't understand. So this diagnostic phase will change as the system evolves, basically. And for right now, it's asking us about episodic memory. Um, so what I'm— and all you need to do to keep continue on is basically just reply as normal and just, uh, you know, type going. And they've also got some different commands here. So if you wanted a hint, if you didn't know the answer, or if you want a status, or you just want to finish the session, then it will go through that and it shows which hints and stuff you've been using. So the first question is asking us about episodic memory. So what is episodic memory? So I'm basically going to say this is previous, previous experience or past event. So this could be This could be a conversation. Sorry, let me see if I can actually type properly. Customer in the past. Let's do that. So what I— although I'm typing here, I also like to use another tool called Whisperflow, which is, um, it allows— it's kind of a dictation tool. So it allows you to press a key on your keyboard and you can just talk back and forth. Um, I like to do that in a, in a in an ideal world. I'm not going to do it today, but because then you can have these long conversations and actually talk to the AI. And then because you're talking, as we mentioned before, when you actually say it in your own words, um, you can— it can kind of understand more what you don't know, what you do know. So now we've actually given it the first answer. It's, it's explained— it's, you know, it's explaining what we've, what we've said and, and kind of teaching us in a way and not letting us— it's not giving the answer straight away. It's basically just asking us to reaffirm what we're saying. And for this one, just for the purposes of a demo, I'm just going to say, I don't understand. Can you explain? And we will see what it brings up. Maybe shouldn't have used the high thinking mode on Claude right now because it's using a bit more tokens, but we're okay. And just to run through, actually, it's done it. So it's kind of given us a, you know, as I said, it's a teaching agent. So it's going back and forth, giving us an understanding of what different semantic memory and episodic memory is. And then it's going to ask us in our own words, can we actually explain that? So you're, as you build using this, this tool, you'll get a lot better understanding of the topics that we're actually, we put into that. So to finish off this part, I'm going to click finish. Actually, I'm just going to run XP debrief on agent memory. So there we go. Can't spell. So after you've had a session, you've now— we've now got this transcript of how— of the conversation we've had, which when we then run the XP debrief, which is another skill that's in here, which I'll show the, uh, the architecture later in a second. Um, this will take everything that we've had in this session and basically understand what you do and don't know and then save this to the database. So this means that, um, the, the first skill isn't doing all of the load. We're actually splitting it out because what I find from using these different AI tools is if you give an agent a massive long-running task, it will really struggle. But if you break those tasks down into really small, uh, sorry, if you break that workflow down into really small tasks, it does a lot more effectively. So in this one, that's why I've broken these out, and it's what it's doing now is like taking what we've had, uh, saving that evidence, and then basically it will save that into the database. Let's see if it runs. I might just carry on to the next slide and we'll see that in a second once it's done. So, I'm going to go back to this. Okay. We'll pick back up on that in a second. Um, so I now wanted to go into the architecture behind it, and because we've explained a few things now, it may be good to see a diagram of what I'm actually talking about. So firstly, we've got the first skill that I've been using is an AI Researcher skill, which was actually the original, uh, idea behind this talk, but I've kind of fleshed it out a bit, a bit more. Um, so the AI Research tool will spawn quite a few different research agents that will search all of the peer-reviewed studies about a certain topic and consolidate that into a research document, which is used as the, basically the source of truth for the XP Builder. So it's not going to kind of, well, it'll limit the amount of hallucinations that it, that the AI does, because if you didn't know, AI can hallucinate, uh, and just basically make up things in one way. So you always have to check, um, what you're using. And, um, yeah, so we've got the AI Researcher skill, uh, then we've gone into the XP Builder, which is what we've just, just gone through as the, as the demo, which is a back and forth conversation where it's probing you, trying to figure out what you, what you learn, what you know, and your while also it's teaching you these different concepts and you're like having this conversation, which I believe is the best way to learn something. And then once you've start to build that up, then you hit the XP debrief, which is the judge of that conversation, and it writes that all to the memory, which as we mentioned, the Unified Memory Core is Oracle AI Database, and that keeps your relational data, your vector embeddings, all in one engine. And in the setup I mentioned about like you can run this on OCI, you can also just run this as a local container instance. So, um, that will be another section. I think I'll update the, uh, the GitHub so you can show how to do that. And then the last part, which we'll show later on, which is kind of the cool bit, which you— the visual cool bit, um, is we've got Obsidian for a front end, which basically Obsidian is a tool which a lot of people have been using recently as like a second brain, which I suppose this is a good example of it. And it allows you to— it allows Claude Code to interact with this, this, I suppose, this framework. So you can house things and also see a nice, lovely projection of your brain. And the last part, which we've not actually gone— we won't go into today, but As part of the roadmap, once you've— you're learning these different concepts, I want to be able to take that learning into creating content, which I'm calling it the Framework Builder, which I'm currently working on at the moment. And what this will do is create a— from what you've been learning, create a skeleton framework of, say, a blog or a, um, or a script for a video or any sort of other piece of content. And once you've got that script, then you can kind of fill in the blanks in a way. So you're also consolidating your own memory. So you're not just like getting AI to just generate it and post it. You're actually consolidating your memory and then you can have this back and forth conversation on how this, how you want to build this out. So that's for a later, later iteration. So let's go to the next page. And this is just a quick slide about what actual Oracle AI Database is and why it's the unified memory core. So it's basically one engine that keeps your facts, which is your relational data and semantic data side by side. And there's no separate databases that you have to keep in sync. So the reason why it's, it's durable is it basically, every concept that you've worked on has its own record. So if you came back 6 months later, and you wanted to see exactly what you understood understand then and what you didn't understand, you could, you could see that it stays there. But then also you can then see how you've changed over time. So you can, you know, you can see your trajectory of how you're learning different topics. And really, the way I see it is like a really good map of your brain, which— bit meta, but anyway, I think it's a cool idea anyway. And then we've got Reason 2 is the facts and meaning are in one engine. So in a lot of AI setups, you actually have to have two different databases. You have to have your relational database, which for an example could be your, all your company records. So your transactional data and employee records and all that kind of stuff. And then you'd have a vector database, which is slightly on the other side, which holds semantic meaning for AI agents. But with Oracle AI Database is all in one, so you don't have to have a few different databases. You have to, you know, keep them all in sync. And yeah, so it's a lot easier. And we use LangChain Oracle DB as the glue between that. So every time I ask the XP debrief to write to the database, it's using the LangChain Oracle DB integration as the glue. Which means you've got fewer moving parts. Um, and the good thing about this whole thing is because you've got one database, it means that, uh, you know, you're from laptop to enterprise, you've got the, the same, the same setup the entire way. So you can scale this really quickly if you want to just start and deploy something quickly and, and have a mess about, then you can actually build that into an enterprise, uh, application. So now we're gonna go into the second demo, which hopefully will be cool. Uh, so this is gonna show SQL and, uh, Obsidian. So I'm gonna try, try and show how the database is actually working. And let me now figure out how to change the screens. I think we are, uh, let's find that one. Apologies everyone, I'm just gonna have to stop the share because it's not coming up. We are in a live demo right now, so that's exactly what's the issue. Let me just close my VS Code and open it back up again. Let's see if this works. And if you do have any questions, uh, I think there my colleagues may be in the chat answering a few. So if you have any questions, we've got a Q&A at the end as well. Okay, I think we've found the right one. There we go. Now I just need to find my other— so I'm just going to pull up this other window because I'm not going to do the SQL commands off the top of my head, and I've got them prepared in the background. Um, so right, let me— let's start with— I think you can see my screen. Yeah, so we've just run that in the terminal. I'm going to close the terminal for now, and I'm going to go to the Oracle SQL Developer. And if I just show on the extensions, I might actually make this a slight bit smaller so you can see. So this is the plugin that I'm using. It allows you to basically interact with your database on the fly. So if you want to run any SQL commands and figure out what's in your database, this is what we're going to use. So let me go back to that and what it looks like just here. So I've already got my, um, let me open this. Let's just see if it will connect. Okay. So I've already got my Limitless Brain, which is in OCI already set up. So I'm just going to run some quick commands to first of all check if we're running.

Okay.

So we're running. So the first one I want to run is this command, which will basically show the latest sessions that we've had within the database. So if I pull this up, so you can see straight away that we've had quite a few different sessions when I've been testing this for the, for the demo. Um, there's, you know, the different topics that we've been using. So I've, uh, there's an Agent Loop topic in there as well, and also the Agent Memory. And you can see the status of it. So it's when we've created a session, we've finished it, we've debriefed, and it's all kept within the database. We can also see different things about— if I run this other— we can also see the different, uh, concept assessments that we've done. So what this all— what it does is to gamify in a way. We've got, um, different scores attributed to, to the different labels and the different topics. So each time you have a session, it has that score of how you, how you're doing before. And if you do well in the session, then you'll increase your score. If you do worse, you'll do worse. So as you— and it also shows like a different band of like how you're doing. Obviously, I've reset this, uh, yesterday. So, uh, I did have a higher score, I promise. Um, and then the last one I want to run through is this one that shows the different evidence for the sessions and, and shows that when we actually run the debrief, um, it's actually keeping what we've said in the database, which means that over time you can kind of see, oh, this is how I thought about this concept before, and now this is how I think about this concept. Uh, the evidence snippet here, I think this is actually the, the latest one we've done. If I look, doesn't give the date, but, um, it basically explains— I think this is the actual prompt, the actual prompt I gave it before, so it kept that in the database. And as you carry on, there's going to be, you know, it's going to have a really good understanding of, as I said, what you know and what you don't know. Uh, that's that section. Now I'm going to split over to the, I suppose, the cool bit, which is the Obsidian frontend, which if I just share here, right? So now actually this looks maybe a bit— we'll do that. I hope everyone can see this right now. I don't know if it's too small, but so this is a work in progress, but this is the Limitless Command Center that I've been building. And what this does is it'll track— it's basically a representation of what we see in the database. So every time the database gets updated, Obsidian basically, uh, we, we reset Obsidian and it shows, um, like an updated view of what you've been learning. So you can see the different topics, a score, which, you know, at the moment it's fragile, which is not the best, but, um, it shows the score, the weakest concepts that you struggled with, on the next review, what you should actually kind of look into and make more concrete. And we've got mainly here the two different topics that I've been doing for the demo. But if you've got any sort of topic that you're wanting to learn, you can use this, this workflow to kind of build that understanding. So this doesn't have to be even be AI, uh, you know, knowledge or anything about around AI. That's just because it's my job and I, I have to kind of know these different things. Um, but if you want, if you want to learn anything, this is the kind of area. And all these will be— it will be populated with the other topics that you've done. Um, shows the weakest, strongest concepts and all that kind of stuff. But there's one section that I really like, is the knowledge graph, which I think I can zoom out there a bit. But what this is doing is it's showing say we've got agent memory here and the agent loop is showing the overall concepts. And then from that, you can see how these different concepts link to each other, which is in my head is such a cool system to actually see the underlying kind of connections between these, these different topics, because obviously agent memory and agent, the agent loop, which if you read the blogs, you'll, you'll kind of understand a bit more. But those topics are very, very closely knit because it's all around the similar topic. So you can see these different bridges. And as you can imagine in the future when you've got loads of different concepts, this knowledge map will be massive and with so many different connections in there. And the reason I like this is because I kind of think it's a, you know, it is in a way a second brain. It kind of knows the relationship between these different, these different areas and how, how you're learning, which is in my head, I, my favorite thing to do is just learn new things. So this for me works well. Um, we've also got in this, it shows each individual concept. So you've got the topic as a whole, which is agent memory. It goes down takes information about the— all the recent evidence from the, uh, session that we've had, um, which, you know, explains the, the topics and how, um, and what we know. Also the relationship, which this is used within the knowledge map, so this is kind of like another visualization of that. And then we've also got the individual concept, so this is where it stores every single concept, the band that you're working on, the score, uh, the trend line. So you could even think in the— as you add different areas to this, you could have different, um, different visualizations other than just a knowledge map to show how, how you're learning different concepts in different ways. And the, uh, it's endless basically. So that's going to be what I'm going to show you for the, the Obsidian. And we'll now just go back to the final section, which will be— if I can set it up correctly and find the screen. Okay, so we've just done this Obsidian walkthrough, and this is where I wanted to just explain a bit more about this like knowledge map and how kind of important it is. So Every concept that you learn becomes a node in that knowledge map that we just saw there. So yeah, every— all these related ideas do connect together, and the more you feed into this database, the clearer of a picture it understands of how you actually think. So it is really looking into being a, uh, a second brain. Um, and as I mentioned, you can add more research documents into here and it will— and it will bring so many different concepts out of those that you can then learn in the future. And this is something that obviously I've been developing and is available now on the Oracle Developer Hub. But this is just the first, first kind of slice of it. The core loop here we've got is, you know, about learning and debriefing and the persistent nature of the memory. But in the future, I've got a few different— I mean, in the, in the immediate future, my next things are firstly to create a framework that builds content out of this, which, um, will then, as I mentioned before, that create that skeleton of then you can use that knowledge that you've built and build out this skeleton, this, uh, blog or, you know, uh, script or anything like that. Um, we could also— another way I actually like to learn with ADHD with someone with ADHD is I use a Pomodoro timer all the time. So I'll do a 30-minute session of a bit of, a bit of work, or 45 minutes or 60 minutes. And I find that, that for me, knowing that I'm just doing that 30 minutes, actually tricks my brain into wanting to do more. So you could do, um, actual time blocking of these sessions. So you say if you've only got 30 minutes to actually do something you can just have a 30-minute session and then you're good. Another quick tip for learning anything, which I found is just crazy, is Google's NotebookLM. And that is a game changer of a tool. You can, within that, you can drop a research document, which, or yeah, a research document. And what it will do is be able to create an audio overview, which is two AIs talking back and forth to each other. And they sound scarily accurate. And they'll basically have a conversation, which is kind of how what I'm trying to emulate here in Tex in a way, where you're having that conversation back and forth and they're kind of giving different opinions and that. And I find that for me anyway, that helps so much in my kind of learning because I'm able to listen to two people talk about one specific subject, say agent memory. Um, and I use that every time I go to the gym or go for a walk or anything like that. I'm just listening to two AIs talk, which is— I don't know how I feel about it, but it's this the age we're in. Um, there is— and then there's also loads of other different ideas that I've got in flight. I'm not going to go into these because there's just— there's too many things that you can, you can build out of this. And we've come to the end. So this kind of topic, or this, this, uh, tool is, is just basically there to help you learn faster. And as I said, this is something I built specifically for me. So if this is not, you know, something you would like to implement, totally understand. I just found that having to learn so many different new topics or, or different concepts It's, uh, you know, I needed a tool and a system to help me do that, and I think this is something I'm going to be building out. So if you would like, uh, I mean, you can message me on LinkedIn, or if you want any, any questions about Agent Memory or Agent Loop or, or whatever, send me a message. And there's also the repo there for the Limitless Workflow, so if you want to kind of, I don't know, try it for yourself. There you go. And yeah, I guess that's it. I appreciate it.

All right, super. Thank you, Cassius. I love that we've built a tool, but then also like there were tons of tech tips on like how do you learn better? Very important stuff. I feel like there's so much going on in the world of AI at the moment, then figuring out how to cram it all into your brain is one of the big challenges. Yeah, so we're going to go to audience questions at the moment. So if you've not written a question yet, everyone in the audience, please, uh, uh, now's the time to ask them. Before we get to that, uh, we've got 3 more sessions, uh, in the Anthropic Age Week. Uh, so tomorrow we're gonna be learning how to build secure agents, uh, using Claude code. So, uh, we're going through a similar kind of idea today but, uh, with a bit more focus on security. Thursday we've got a panel interview on, uh, how do you adopt Claude within the enterprise. Lots of challenges there. And then on Friday, there's going to be a session getting into more depth in Claude Skills. Now, after that, we're moving slightly away from Claude, but we've got a 2-week webinar series on agents in the enterprise. So, uh, basically lots more stuff on agents. I mean, there's an unlimited amount of stuff to talk about agents, I think, at the moment. So, uh, if you've not signed up for that yet, uh, there is a link in the chat. Please do join us for those. Um, All right, so, uh, I think there were some things, uh, in the chat, people getting a bit confused about all the different moving parts. So, um, we took— we made use of Cloud Code, Cloud Skills, uh, there was, uh, the, uh, Oracle database, uh, and then, uh, it was Obsidian, all different bits of it. Do you just want to give us a recap of all the different bits of tech we used and what they do and how they fit together?

Yeah, definitely. So, uh, yeah, it was pretty confusing now I think about it, um, but Take away the Obsidian part, that's just an optional front end that you can look at and see things. Really all it is, if you clone this repo, you can actually get your Oracle database set up. It's basically just those two things, Oracle AI database and Claude Code. And within Claude Code, I'm using that as the interface. So you don't really have to go outside of that. You can just have these conversations with Claude Code. And it will basically save that to memory and you're continuing to build this kind of second brain out of it. The Obsidian part is just, I thought it would look cool and it's, it's, it is cool. Yeah.

Nice. Yeah, certainly. And yeah, Claude Code is just like, I like this, just an interface. You just chat to it and it's writing the code for you and that's just how you get work done.

Oh, there was another thing just to mention as well. This entire doc, I don't know if you've seen the Claude design came out recently, like a few days ago, and I basically created this entire doc using AI as well, which is crazy in a way because like the, the me creating slide decks on PowerPoint is just out of the— we're not— I'm not gonna be doing that anymore. Um, and Claude Design is a really interesting tool where you can, uh, create amazing looking things. You could also do this— I actually ran out of, uh, my Claude design kind of the amount of tokens. So I just imported this into Claude Code and you can just carry on building things like a slide deck like this with Claude Code. You can honestly do so many things. It's actually overwhelming how many things you can do with Claude. So if you've got any questions about that, you can message me on LinkedIn.

Very versatile tool. I think, yeah, the days of building slides in PowerPoint and Google Slides, like I don't think anyone would be sad if that's over. It's like, yeah. Not often a fun task. All right. So, oh, one more question before we go to the things from the audience. There was some discussions on like, when do you want to have the memory for your agents? Just, I guess, within the, within context, like within variables in your program or in flat files, or when do you want to move to a database? Like, how do you decide whereabouts you want to, you want the agent to remember stuff?

Yeah, this is a good question. Like there's the whole kind of, I see it as like an illusion of memory when you think about the context window. So Context window is kind of like when you've got a chat with Claude Code or you've got a chat with any other kind of AI tool. So ChatGPT, Gemini. And when you're having that specific chat, it kind of is remembering everything you're talking about while you're having this conversation. It's remembering files that you've given it and everything like that. But that context window is kind of self-contained. And then when you start a new chat, then it hasn't remembered anything that you've, you've talked about in that. I mean, Claude does have a memory feature which kind of takes a few different sprinkles of what you, you're, what you'll, um, say in the chat and save that to a memory. But along the, um, on the way, you, you will miss out on so many different things that you, that you want. And if you're having to say create agents or like a multi-agent pipeline, you can't rely on it just being in the context window. You have to have like a, a, a store of that memory. And I, I believe the best kind of store for that is a database.

Okay, yeah, certainly. I— once you're— if you want persistence, you want like your agent to remember things between sessions, then yeah, it doesn't need to be in a file. And like, certainly like a database makes a lot more sense for structuring things. And once things start getting a bit bigger Uh, all right, let's go to some audience questions. So, uh, Fossil was asking, uh, can you do the same thing with the skill on Google Antigravity? So Antigravity is like the, I guess the equivalent of Claude skill. It's like a Google Development Platform, right? Uh, I'll let you explain.

Yeah. So, um, skills are not specific to Claude. Um, they're basically, I mean, I'd say industry-wide now you can take a skill. I use, um, another which, uh, shout out my friend Nacho who showed me this, uh, py.dev, which is another kind of like, uh, terminal similar to, similar to Claude. And the skills that I create for that work in Claude, they work anywhere. All the skill is is just a markdown file, which is like, you know, used everywhere basically. It's just, it's just a file with, uh, a name, description, and some sort of, uh, task that you want it to do. So yeah, you can use it in any sort of whatever your AI of choice is. If you want to use Antigravity or Google CLI or any of these different things, you can basically do the exact same thing. So obviously we're talking about Claude today, but it's very— yeah, reproducible. I don't know if that's the word, but anyway.

Portable, maybe.

Portable. That'll work.

Yeah. And just massively cut down the amount of boilerplate you're writing in your prompts because it's like, well, it knows all that stuff it's containing the skill. So yeah, save you a lot of typing. All right. So next question from Camille, can Claude Code interact with local command line tools?

Yeah, it's Claude Code is it can, what can it not do? But no, it can definitely interact with so many different tools. So like when I was mentioning before about what an agent is. An agent is kind of just, it's a large language model, which is your, you know, basically trained on a large amount of data from the internet and stuff like that. And it's contained within a harness which allows the large language model to connect to different tools. It can search the internet, can read your files, can run scripts. It can do most, a lot of things, which in one way is amazing, but you also have to have safeguards behind it. You don't just want Claude or anything to just have access to everything. There's whole things about creating sandboxes and stuff like that so it only can have specific things that it accesses. But yeah, overall, Claude can connect to a lot of different things. Yeah.

So I'm all right thinking basically as long as you can set up a model context protocol interface, basically it can connect to any tool.

Yeah, yeah, yeah. Most, um, yeah, exactly that. Okay.

All right. Uh, nice. So we'll be covering, uh, this next one, uh, this question from Racy, uh, more on Thursday. But, um, if you're trying to build sort of enterprise-grade products, how do you scale stuff? And has it got all the other sort of enterprise requirements? Like, what do you need to care about?

Yeah, definitely. So, I mean, that's one of the kind of, uh, amazing things about, like, Oracle AI Database is the fact that all of your data is in one area. You're not having to, like, you're not having to have, like, 4 different databases— sorry, 4 different databases for your, like, um, your employee records, you're caching different things, and it's all in one. So it allows you to scale in such an easier way. Um, you've also got different things like JSON-Duality in there as well. So you've got so many different, uh, forms of data that is in one database And that is where I think Oracle, because it's kind of known as a secure place to house your data, that allows you to go from developing something just in your spare time to then actually building this into something that all of your agents can actually interact with in a kind of safer way, if that makes sense.

Okay, so yeah, if you're going to do enterprise products, you probably want like an enterprise data store. So yeah, Oracle is one of these sort of enterprise databases. All right. Nice. So, oh, I think cost is top of mind for a lot of people. So how many credits do you need to do all this? And yeah, just give us a rough idea of like pricing for building these sort of agents.

I mean, to be honest, like I, I do burn a lot of tokens to be fair, but you could do them. You could actually do this with, you know, the normal Claude subscription, the like $15 or whatever it is. Like, I— you could do all that just, just within that, um, you know, your allowance there. Um, but it just depends on how, like, how you want to build these applications out and how much you want to. I mean, there's also the fact that the more, um, skills that you add and the more, like, the more stuff you add to the, the context window, it's actually having to use more tokens to to build these different applications. So you want to— there's a thing called context management, I believe, if I'm off the top of my head, which basically is about efficiently giving the LLMs or Claude Code just what it needs to do the job. So you can be— there's ways to be a lot more efficient about how you use tokens and also which lowers the cost. So there's so many different kind of tools that will help you with that. I'll see if I can put some in the, in the doc afterwards.

Yeah, definitely. So, yeah, have a look around for tools for managing your costs. Be sort of mindful about what you're spending unless, obviously, unless you're spending some alternative money. Nice. So, we've got a couple of minutes left. There's one last question. Now, I just want to say, as part of that Agents of the Enterprise campaign, we've got an Open Claw session on Monday. If you're interested in Open Claw, please do come back Monday. But yeah, can we do any of this with OpenClaw and these other tools that sort of connect lots of other tools together?

Yeah, I mean, these tools have got so many different applications for them. You could definitely, I'm sure you could definitely do that. I've not messed around too much with OpenClaw, so I wouldn't, you know, I'm not going to give my perspective on it. But a lot of these different agent frameworks can do similar things in a way. So it's like, you've got to think of these things as like, You know, people say, what's the best tool to do this? What's best tool for that? You just think of it as a tool and you, if you use it in the right way, you can, uh, you can build amazing things. But if you don't, then, you know, you're going to have, um, I say issues, but you're going to waste a load of tokens doing, doing the kind of stuff you want to do. So short answer, yeah, but I can't give too much else other than that.

Yeah. So basically, uh, there's so many different tools, you can build things in so many different ways. You just need to be mindful about are you building the right thing and is it the most efficient sort of toolset? So yeah, tricky question, a little bit open-ended, but thank you very much, Cassius. That was a very cool thing. I love that, like, build something to help people learn. It's something very dear to our hearts here at DataCamp. All right, so yeah, thank you, Cassius. Thank you to everyone who's interested in the question. Thank you to everyone who showed up today. Please do come back tomorrow. So we're building more things with Claude Code, this time around building secure agents. All right, see you all again soon.