985: The Four Types of Memory Every AI Agent Needs, with Richmond Alake

22:15They're very popular now at the time of this recording. And skills.md file are basically marked down that covers instructions on how to do certain tasks that are given to an agent. I call it the SOPs for agents, essentially. Standard operating procedures. Exactly. But these are basically procedural memory because humans have the same concept of just routines and skills, we store them in a particular part of our brain. You're the neuroscientist there, so you can fact check me on this. We have this part of our brain called a cerebellum. And that stores all our routines and skills. So you can actually think about that in the same context of agents as well.

22:56We are writing a bunch of skills as an MD file and we're storing them in people who are using file systems. And we could talk about file systems versus databases for agent memory. But over in Oracle, we're experimenting with putting these skills into the database, right? And actually putting them within tables that have some different representation of this data. One of them would be a vector representation. And then we can progressively expose the skills at the right time. So rather than giving the agent 100 skills or MD file, within a context, we can just retrieve the actual skills that we need at the time, which allows you to start to scale.

23:36essentially. But that's an example of procedural. Do you want to do that?

23:39Jon Krohn:Yeah, I'll try to come up with an example here. So if you think of some kind of thing that you do regularly, like making scrambled eggs, it's not that complicated. And you probably don't remember the first time somebody showed you how to scramble eggs. It's not like an episode that you're like, oh yeah. I mean, maybe. There might be some people who are like, I love the time my grandma showed me how to scramble eggs. You might have the kind of episode that you can recall. But for most people, it would just be kind of something you can make scrambled eggs now, and you don't really remember how it happened, but you've just kind of figured out through experimentation, maybe some things you read online or some things people taught you, a cookbook.

24:15Jon Krohn:There could be all these different inputs that over time created this procedure that you're aware of. You're not really, there's not any specific episodes tied to that procedure. You just know how to scramble eggs. Exactly. Exactly. That's a very good example. So let's go to semantic. Semantic can be a bit easier. So semantic is actually, just the easiest one would be world knowledge right you can have knowledge about a certain topic that's semantic memory and for agents that could be all of the let's say the institution the institutional knowledge within your enterprise data right how to do certain things uh some abbreviations or some words and you could just bring that into your agents to get them to actually start to have the knowledge that the same um your employees have in terms of that particular use case so that's semantic memory there are different types but that's the one we're gonna use today so over to you yeah exactly so this is kind of like you know any kind of knowledge that's in like an encyclopedia like wikipedia these kinds of key pieces of information again so it's different from procedural in that it's not necessarily a sequence um it's just it's a fact it's an association exactly um you know richmond lives in london it's like i know that semantic fact It's not associated with a specific episode.

25:32Jon Krohn:It's not associated with a procedure. It's just a piece of information that I know. Exactly. And then we can talk about working memory, which is another type of memory, right? Working memory is this short-term, specifically a subset of short-term memory. Working memory is what you're using in real time, in the context. There's a bunch of information that is my working memory that I'm using to speak to you right now to actually interact with you. I'm not having to think for an extensive period of time, right? So that's my working memory. The best way I would describe this in within a genetic context is the context window of the LLM.

26:05That's working memory. And over to you.

26:08Jon Krohn:Yeah, now that makes perfect sense in the context of an LLM. You know, it's kind of the, when you think about it from a computing perspective, this, you're not going to need to do some kind of retrieval. You're not going to need to look over a database or do some kind of rag process in order to find whatever's in the working memory. It's in context in the agent. And this is analogous to the same experience that we have when you stay focused on some complex task. Like right now, I'm trying to find as much time as I can in a given week to be writing my next book, which has been difficult, difficult thing to find time for.

26:45Jon Krohn:But it's really important that I carve out time to be focused on book writing for several hours at a time, because it It can take me up to an hour to kind of get all of the relevant context back into my working memory in order to make progress for hour two, hour three on the book. Because there's so many, you know, where are we in the book? What do readers already know? Where are we going next to get all that kind of stuff just in my working memory? It takes a bunch of time. And yeah, it's also for agents, it ends up being one of the key, one of the key pieces of memory to have in place. Yes. This is good.

27:19I like the back and forth. Very educational. But one thing your listeners can actually get is we cover all of this and more in a recent course we just did in partnership with Deep Learning AI. So Oracle and Deep Learning AI, along with Andrew Ng, we just released a course on building memory aware agents.

27:38Jon Krohn:Yeah, that's so cool. I saw on LinkedIn when you guys released that course, you had the announcement, you and Andrew Ng together. That's pretty cool. Yeah, and a couple of my teammates as well, Nacho. And this is something that is very interesting about Oracle. And I joined Oracle, it's been about five to six months now, is there are so many people that understand the agent memory application and how useful they can be in actually building AI application. And when you look at Oracle, Oracle has existed for over four decades and it started off as one of the oldest database companies. So I should not be surprised that they understand how things are shifting in the paradigm of change itself.

28:18So we were very quick to actually take on this agent memory narrative and start to meet developers where they are. So the course we did with Andrew answers a lot of the questions we saw developers asking and our customers asking, and what we think are the patterns that will begin to solidify themselves to what might look like a standard, which I'll obviously like to contribute to. Of course. But also, we recently announced that we're going to be putting out an agent memory SDK in Python for developers that like to get a bit hands-on. And we also have a private agent factory that allows people to actually build agents using no-code tools.

28:57So Oracle has a massive understanding of the shift that is happening and how to actually bring these to developers and our customers.

29:05Jon Krohn:Fantastic. And we'll make sure to include links to all of the items that you just mentioned in the show notes for our audience members to be able to enjoy. It sounds like you get to have a lot of fun working with clever people at Oracle, creating useful materials for our listeners. Yeah. To dig into a buzzword that I said not too long ago, I mentioned RAG. And you actually, you'd noted here how agent memory goes beyond retrieval augmented generation. Where do you see RAG falling short for agents? Well, RAG is retrieval, essentially. It's just retrieving information and parsing that alongside with the user prompt or user objective to actually augment the generation of your LLMs, right?

29:52It's a technique that's very useful to actually ground the LLM in domain-specific data. Memory, you do need to retrieve memory, but you also need to update memory. You need to consolidate memory. You need to resolve conflicts between memory. You actually need to forget memory. React doesn't cover all of that, right?

30:11Jon Krohn:So, clearly. A good example of where it could break down is my previous startup that I worked, that I was a co-founder at until about 18 months ago, we were specifically building applications in the human resources space. And you could imagine, it's kind of easy to think, okay, if I want to have this human resources agent working, let's just throw all of the HR manuals that we have in the business into the memory and do reg over that. But the conflicts are a big issue because there can be directly conflicting information in these HR manuals. One HR manager creates one in 2023, and then another one creates one in 24.

30:49Jon Krohn:And how you're supposed to resolve some specific kind of HR incident, there could be directly opposing information. And so that is one of the kinds of situations where reg can fall short because you can retrieve one piece of information, you know, the 2023 information and tell a person to do it one way on the first day, and the second day come back and tell them to do it another way. In both cases, they were retrieving, as far as they know, correct information, but conflicting. Exactly. RAG is a useful part within the toolbox of agent memory, but there is so many more, which again, we cover in the course, we cover in our technical assets, we cover in different blogs that we put out as well in our Oracle developer resource.

31:31Nice.

31:32Jon Krohn:Is part of it to have in the working memory some kind of consolidation that resolves these ambiguities and these conflicts? Okay, we can get into the technicalities. I wouldn't have that in a working memory because actually resolving some of the technicalities and conflicts can be computationally expensive. Right. And you probably want it to be permanent. Yeah, when you want a story, you want it to be permanent. You also want to think about latency as well, right? You don't want your user waiting for your system to figure out which memory to use and how to resolve conflicts. That's a bad experience.

32:07This is why memory engineering is such a unique, not unique, but it's an important discipline. But again, nothing I've said is new. Latency and the reduction of latency and retrieval pipelines or information system, we've been trying to solve that for decades within the database world. So there are techniques like quantization, right? reducing if you're using vector embeddings you can reduce the precision data types of your vector embeddings to speed up speed up the retrieval or reduce the storage space utilization so these are things that might get lost in people that are just coming into the space right if you're vibe coding an ai application today you are not going to know about quantization right so we need to actually start to understand these and start to start to educate the new people coming into the space but as well as bring the people that have been here for decades along with us.

32:59For sure.

33:00Jon Krohn:I think in this vibe coding paradigm, it is still valuable for the human vibe coder to have some understanding of how these systems work and how they could be improved. Maybe it will even be important for people to still understand the Hubel and Wiesel stuff that we were talking about earlier in this episode in terms of being able to understand what's possible because yeah, you can end up, I think probably for years to come, I could end up being wrong. It's such a fast moving space, but there's a broad context and an understanding of where a product or an experience could go to that the LLM is not necessarily going to be able to come to a conclusion on its own.

33:44Jon Krohn:So yeah, get out there and learn from Richmond's courses. Yeah, and one of the things that actually sort of like dawned on me when I was speaking to the developers and startup folks working on agent memory in New York this morning was they shared with me that they don't see a general purpose solution to memory because memory is so unique to different use cases and different niches and different workflows, right? That they think there might be different, that the ideal solution would be different ways of handling memory will probably emerge. And I don't think there'll be one general solution, which for me, standing from an infrastructure point within the stack, I'm thinking more of the general application of the solutions that I'm creating, right?

34:32I want to solve memory for all. And today I was hearing, well, it might be a bit harder. How about we solve memory for this specific use case and then we just keep tackling things on. And that was a very interesting insight. And it kind of shaped how I think I'm going to be looking at solutions and educating developers. And you'll probably see more of these in the next coming week for the listeners in the events that we're going to be in. So we're going to be one of the biggest events, which I'm sure a lot of AI folks go to, is AI Engineers in San Francisco. So we're going to be there. It's June and July this year.

35:09So we're going to be there. So if you see any Oracle folks, just say the two words, agent memory to them and they'll go off.

35:16Jon Krohn:Nice. I like that. And it sounds like you also have, we're expecting this episode to be out April 21st. It sounds like you have an event next week at time of release. Yeah. So we do. The one thing that we are trying to do over here in Oracle AI database is meet developers where they are, right? So we are meeting them in terms of where they're actually educated where they get information from, which is why we partnered with Deep Learning AI. We're in some of these critical events as well, AI engineers. And one of those events is AI Developer Conference in San Francisco, also by Andrew Eng and Deep Learning AI.

35:52So our aim is to just serve developers and meet them where they are.

35:56Jon Krohn:Fantastic. Yeah, lots of great events for listeners to check out in person. Kind of back to the technical stuff around agents. Let's broaden out a little bit from just agent memory to the broader stack where memory is a key component of that. There's a lot of noise around agent frameworks, orchestration layers, tooling. How do you, Richmond, think about the agent stack today? And beyond memory, what are the layers that really matter? So the layers in the agent stack, right? So one thing is very clear is the agent stack is expanding, but also at the same time, it's consolidating. But the layers that really matter, we have the orchestration layer, we have the connectivity layer, we have security, right?

36:43That's definitely an important bit. We have the reasoning layer, we have the memory core, which is what would be where you're storing your data. We also have the memory managers as well within this agent stack. And I'm sure I'm missing some layers, but the key thing I'm seeing, right, with all of this orchestration, reasoning, memory managers, the way the field is now defining this whole space or how they're actually implementing the agent stack is they're using the word agent harness to describe bringing all of these layers and tool sets within this layer together into a solution.

37:15Jon Krohn:That is something really recent, isn't it? That everyone seems to be calling it an agent harness. It's recent and it's popular because people were saying it maybe about a year ago, we're in April now, right? Yeah, well, yeah. Memory, exactly. We're in April now. So I think Anthropic put out a piece last year, May on effective agent harness or effective harness, right? People have been talking about building systems that work together to create this outcome of maybe some form of agent execution that you want. We all focus on coding agents, right? So if you use cloud code or use codecs, those are agent harness around the specific models to actually allow for a specific outcome.

38:02So this is what the field is defining all of the key parts of the stacks working together. And I think understanding your agent harness is important. Understanding how you can customize your agent harness is important. And obviously, if you're talking to me, I'll tell you to go build a memory first agent harness. where you're thinking, where you're approaching the implementation of your agentic system by just thinking the information that comes into this system needs to be recalled and forgotten. If you just take that mindset shift, it just changes the way you approach how you model your data, the selection of tools as well.

38:43For example, you start to think, okay, if I'm going to be storing information, I probably don't want to be using multiple databases. right you probably want to be using a database that can handle all of the different heterogeneous nature of data that you're going to see in production you can use multiple databases i've seen that anti-pattern happen in in some of the ai teams i talk to but i explain it as this our job is to reduce the cognitive load for llms right and when i say our job i mean the developers right we're meant to give them all of this context so that the llms could focus more on the reasoning that they have to do.

39:22Well, if you take that out of the context window, if you add more infrastructure, more components to your infrastructure and complicate it, you increase the cognitive load for not just the LLM because now the LLM has to think about or have an awareness of multiple different databases. You're also doing it for your developers. Because now your developers have to know how to work with multiple databases and all the code that glue them together. So reduction of cognitive load is this thing that I'm really trying to emphasize. Because in AI, we find that speed of experimentation, of implementation, it's a moat, right?

40:03You have to be in the arena. You have to experiment. If you're thinking about how you're going to glue database one with the next database and what database should I use or query languages, then you're wasting time. The AI train is moving.

40:16Jon Krohn:It's moving fast. So what do you think, for our listeners, what do you think are the parts that they should be building themselves in relation to agents versus what they should be taking off the shelf? Yeah. So the one thing that is very important is understanding the domain that you're trying to build for, right? Workflow expertise is so important. It's so crucial. There is no way I can build a healthcare agent better than an AI engineer that works at maybe one of the large healthcare providers. Because they have that domain expertise of everything that goes in there. I can give them some building blocks to help them think about it, but there's no way we can compete.

41:00Same thing with legal. So one thing that the engineers today should be focused on is understanding the workflow you're trying to automate or you're trying to optimize. Understand the workflow in detail. Then the one thing that you probably shouldn't be wasting, not wasting time on because tool selection is important. And I do think looking for a general purpose database, one like Oracle AI database that can handle all of your different data types. So using the Oracle database, you can handle vector data. You can handle graph data because everyone likes to do graph rag or knowledge graphs.

41:36Jon Krohn:Especially for semantic knowledge. Exactly. Especially for semantic knowledge. And you want a database that can handle all of the relational data as well that you normally have and lexical and spatial, right? All of these different data types. So we can handle all of that in Oracle AI database. And we bring in a lot of what developers have to think about building this AI system. So within the Oracle AI database, you can have embedded models in the database, right? So that reduces some of the cognitive load on your developers to start thinking about, okay, how do I start building this ingestion pipeline with the right embedded models?

42:13We are looking to take away all of this concern so that our customers, our developers can really play to the advantage of speed of innovation.

42:22Jon Krohn:Nice. Thanks for breaking that down for us. of all of the types of agent capabilities that are being incorporated into these agent harnesses. So things like planning, tool use, orchestration. Do you think that memory has been under-invested in so far? I think there is a focus on top-in benchmarks from the research labs, right? They're very benchmark focused. There is also a focus on the particular coding agent use case, right? I do think memory was ignored initially because, and I say this because two years ago when I was seeing definitions of agents, memory was always missing. There was tools, there was reasoning, there was planning, but there was no one mentioned memory.

43:07And I was like, wait, that's the most obvious component of an AI agent. I think memory came a bit later, but now everyone is paying attention everyone's realizing that we need to solve the ability for this models to learn right so there's agent memory there's another paradigm and again nothing i'm saying is new there is the whole school of continuous learning right that's existed probably in parallel with ai essentially right so nothing is new it's just bringing what's old and just surfacing it up so we can bring the right attention to it so we can solve it.

43:44Jon Krohn:Yeah, I think some people were grouping memory in as another tool. But I think that's too, it doesn't give memory the importance that it deserves in so many agentic applications. Yeah, and I'm biased. I call it the final battleground, right? Because I think if you solve memory in a very meaningful way, just hearing what I heard this morning speaking to the developers, they're just waiting for a solution that just handles the whole retrieval of information and forgetting of information and doing it very, very well in a way that the system is just reliable. You can just, because you don't have to think about the way you remember things, right?

44:25As a human, you don't have to think about it, right? So developers are looking for that. So I do think they're expecting it to come from the research labs in a sense, but they're also expecting it to come from the database company. This is why I'm talking to them. But I think it can come from anywhere in the agent stack. And agent memory is one of the fields where you can have a single developer create a solution that could probably compete with one that comes from a researcher.

44:52Jon Krohn:Sure. Maybe because it's so easy today. That's why we're seeing so much fragmentation. Do you think that the agent stack fragmentation is here to stay? Or do you think in the future we might consolidate in the same way that the web did? Yeah, so is the agent stack going to consolidate? You mean by just in terms of tool selection, in terms of players? Yeah, so historically for the web stack, there were lots of different ways of doing things, but more or less today people do it all the same way. Yeah, back when I used to be a web developer, the biggest battle was are you going to use Angular or React?

45:36right that's what that's what kept people off at night then vjs came and wherever um but i do think that developers like to experiment with new tool sets that come out but i do think that there will be some staple in your stack that you're just going to use whenever you want to use your models as in model the the key model players now in the stack are there quite few you don't have loads except if you go into the open source direction i do think there's going to be some consolidation and not because and that's going to be because the tool selection is not where the fun is right it's every order value is what you're actually building and getting to your consumers and at some point we're going to have key players like oracle that are very in the in the problem that we we've caught for ourselves which is agent memory we're going to we're putting a lot of resources to solve that to solving that and then when we do using our solutions in wherever of form factors they come in will get you to where you need to be.

46:37So you don't need to essentially shop around.

46:39Jon Krohn:Yeah. Yeah. So in the same way that Python became a de facto standard for doing AI backend programming, some aspects of the agent stack will emerge as de facto standards. You somehow just pissed off a bunch of Java developers. Do you think there's Java, well, I guess so. Yeah, it's Spring AI and new existing. I don't know how many listeners we have that are developing their AI applications in Java, but I'm sure there are great advantages to going down that route. Someone will have to come on, reach out to me if you are in that situation and we'll have you on to explain why. Nice. All right. So we've talked a lot about agents.

47:24Jon Krohn:Let's talk about the specific solutions that you're kind of representing here. So you lead AI developer experience at Oracle, specifically around the Oracle AI database. You've mentioned already earlier in this episode some of the key reasons why you think the Oracle AI database is a great solution for people building agents, things like being able to handle reg and graph databases all in one place. What are the other reasons why the Oracle AI database is a compelling platform? So one thing, and this is important actually, and I'm speaking as an AI developer here, you probably notice AI developers don't care about security as much.

48:09Jon Krohn:It is a personal problem I have. It's kind of an afterthought for me, which is silly. It's because we have to experiment, right? We always have to move faster after experiment. look at the uptake of OpenClaw or even MCPs, right? One of the key advantages of Oracle AI databases, all of that concerns is taken care of. The Oracle databases existed for decades. We've had to really understand the nature of security and privacy in different regions with different regulation, with different compliance, and we've been meeting them for decades. So again, AI developers don't care. They're like, okay, yeah, we don't care.

48:49How does this help me build AGI, right? How does it help me make my agent be more autonomous, right? The key thing that I tell developers is security is important. Privacy is important. When you start to deploy most of the systems in some industries that are regulated, like the finance industry, the healthcare industries, the compliance you have to meet that every single component of your AI infrastructure has to meet. So using one that actually just takes care of it allows you to do what I said is important, which is move to market much quicker. So the Oracle AI database, we take security and privacy as a top priority for us.

49:28And we remove that cognitive load, again, from our developers.

49:31Jon Krohn:Cool to hear about the Oracle AI database. With your role as leading AI developer experience at Oracle, can you tell us a bit more about what Oracle is doing to allow people who are building to benefit from what Oracle is doing? Yeah. So Oracle is actually, like I said, meeting developers where they are, meeting builders where they are. And one thing that we're doing is putting out a bunch of products and tools and features that actually allow developers to build quickly and at the speed of innovation. So one of the things we released quite recently, and I think I've mentioned it already, is the Oracle Autonomous Vector Database.

50:14And this is a vector database solution that Oracle is putting out there, currently in private preview, and developers can use it for free today. You have a solution where you can store your vector data for free. But also, Oracle is actually in the space of actually providing some tools, some open source tools that some of our developers can use for building agentic solutions. So one of them is AgentSpec, which is this solution to actually making agents portable across different frameworks. Because one thing we saw was this sprawl of people building their agents using different frameworks like Cray-AI or using LankChain or using Lama Index or A-Stack.

50:58And we thought to ourselves, we need to standardize this and consolidate it because in enterprise, we actually have to have a common way of talking between large organizations. And the team put together a open source tool called AgentSpec. And behind that is another framework called Waveflow. So we are definitely trying to meet developers where they are, either by giving them some of the free resources that we provide over Oracle, or even just going in there and putting all the learnings we're showing in open source tools.

51:30Jon Krohn:That makes a huge amount of sense. It sounds like a great reason to go with the Oracle AI database. Obviously, we've talked about the deeplearning.ai course that you co-instructed on agent memory. You also run live trainings in O 'Reilly. I think I might have mentioned that earlier in the episode, but can't remember for sure. So people who have O 'Reilly subscriptions through their work or school or personally, you can look out for Richmond's upcoming trainings. Let me actually talk about that. Yeah. So I started that series with O 'Reilly as just the main mission, which is educate developers about agent memory.

52:06And I started this around this time last year, right? Not a lot of people were talking about agent memory and I thought it was very important. And I thought, okay, let me actually start to educate people on this. And O 'Reilly, they said it was a good idea. They've seen some early signals. We started off with two hours, right? And we got a very good turnout. So we did about four iterations of that and increased the time because there was just more, more. And now it's turned into a two-day boot camp, right? With O 'Reilly. What started off as two hours in a year is now a two-day boot camp with five hours on each side.

52:40Wow. Talking about age of memory. Oh my goodness. And the next session is going to be in June 16 and 17. So the developers that we're seeing in the future are going to be very important because they're going to start to, for lack of better words, they're going to be the architects of intelligence. Right. There's a book called The Architects of Intelligence that people could go check it out. That book is actually quite interesting. I'm going off tangent here. That book contains a lot of predictions of things that are going to happen in AI that I think 95 % were wrong from really credible people.

53:13It's called Architects of Intelligence by Martin Ford, I think. It's a good book.

53:20Jon Krohn:You're saying the predictions were 95 % correct or 95 % incorrect? No, wrong. Incorrect. Incorrect. But you still think it's a good book. I think it's a good book to see how the, like, we had people like Andrew there. We had folks from Google interviewed. it was like an interview book and you're like I think Demis is also interviewed for that book as well it's a good book because you get to see the minds of the greats within AI really flesh out how they how they're trying to predict what happens but we're looking at it now and we see oh even this people that are in the arena don't even know but it's a good book because you get to one give yourself some grace because if they can't predict it.

54:01Right. Yeah, it's true.

54:04Jon Krohn:You know, I do go around with kind of a lot of guilt of, you know, why can't I predict what's going to happen in AI better given this is what I'm doing all day. And also another big guilt thing that I have, and this is actually going to lead into my next question. Another big guilt thing that I have is that I always feel like I don't know enough. And so for those of our listeners who feel like they don't know enough about agents, maybe even particularly agent memory. Do you have any particular resources that you recommend that they dig into to get started? Yeah, so one good spot, very easy, is the deep learning AI course we did.

54:42Memory-aware agents, it covers everything within a good time. But also you can go to our Oracle Developer Center. The team is always putting out content. And then recently we put out a content called File System versus Database for agent memory. It was a good conversation a couple of months ago and we put out some technical contents on there. So I mentioned three parts where you can get things. But I want to talk to the feeling that it's very hard to keep up in this AI space. Everyone's feeling that. There is someone called Fei-Fei Li. She's known as the godmother of AI. And I was watching an interview at a conference And she said herself, it's so hard for her to keep up with the space.

55:30And she's the godmother of AI. And I think things move so fast. And if anyone tells you they're an AI expert, they're probably trying to sell you something. But I would class myself as a lifelong learner in the space, right? That is what I signed up for. and it shows in how I'm teaching folks. I'm just learning and I'm teaching and I'm building and I'm just repeating that over and over again. This is our third episode, right? First one was over in computer vision and deep learning and it's the same cycle, just being in the arena, learning, building, educating. And that's what we're doing over in our course.

56:15Jon Krohn:Yeah, you're always evolving and growing. That's for sure, Richmond. Do you think that that Architects of Intelligence book is your book recommendation for my listeners today? Oh, we're doing book recommendations. Yeah, we always have that. It's your third time here. You don't remember? Yeah. But you can remember. You already have that one, Architects of Intelligence. You can just lean on that. Yeah, it's a good book. I want to lean on something better, right? There is a book called Thinking in Systems, and it's made by an author called Donello. I do think, and it's not a technical book, I'll say it out loud right now.

56:49And I think it's important that us technical folks start reading non-technical books. And the book touches on how humans think in two systems, system one thinking, system two thinking. And one of them is the fast thinking and the other one is the slow.

57:02Jon Krohn:Yeah, system one is the fast one. System two is slow. Exactly. So, and I think I was reading that book and I think it applies so well to what we're doing in terms of this agent engineering and building memory systems, right? I do think, and I don't want to overuse the parallelism with human memory, but I do think that we can learn a lot from ourselves because nature is the best architect of intelligence, right? So we can learn a lot from the way we process information. Another book that is, so that's my book recommendation. Okay. It's Thinking and Systems, but another one that I think goes with it is, I've not read it yet, but it's literally next on the list, is The Organized Mind, right?

57:43Oh, yes. Yeah, have you read that?

57:46Jon Krohn:It's actually sitting, so in my like childhood bedroom, it's a house that my parents still have. Yeah. It has been sitting on the bedside table for years. And I'm always like, oh man, I can't wait to get to the organized mind. So yeah, Thinking in Systems and the Organized Minds are two books. I don't know if you gave me the recommendation, but I was trying to come away from the non-technical side because I do that all day, every day. and just trying to go more into abstract neuroscience, obviously, and just draw inspirations from there. Yeah, I mean, it's interesting. I mean, I would say that both of those books are still technical.

58:21Jon Krohn:It's just that they're not software specific. Yeah, they're not software, yeah. Yeah, yeah. It's not like you gave us a novel as a recommendation, which lots of our guests do. Really? I'm boring. Next time. Nice. All right. And then this has been an amazing episode of Richmond. As it always is with you, I enjoy our time together. I always learn so much. You've already given us tons of resources for learning more about agents and agent memory after this episode. What are also ways that people can follow you or Oracle on social media to get more information as it comes out after this episode? So to follow me, just follow me on LinkedIn, Richmond Alake, and you'll find me also on Twitter or X, still experimenting with it.

59:07But LinkedIn is the best way to see what I'm talking about. And for Oracle, we have loads of, Oracle is on all the social platforms. So Oracle developers on LinkedIn and also Oracle developer on X as well and on Instagram as well. So we are everywhere trying to educate, build, and yeah, we're in the arena.

59:25Jon Krohn:Nice. Well, we'll be sure to have links to all of those in the show notes. And thank you so much for taking the time out of your time in London or every time in New York to join me here. Wow, that's a weird memory glitch or some kind of cognitive issue. There's a memory conflict right there. Need a brain scan. Yeah, fantastic. Thank you so much for taking all this time with us, Richmond, and yeah, educating us all on all these great things, Asian memory related. Hopefully we'll be able to have you on the show again sometime soon. Yeah, I'm looking forward to the next city. We did one online. Tokyo.

59:58We did one in London. We did one in New York. We're going to do one in Tokyo. Let's do that.

1:00:03Jon Krohn:Yeah, let's do it. Thanks for having me. Perfect. Cheers.

1:00:08Jon Krohn:Always love having Richmond Alake on the show. This time around, he magnificently covered how agent memory is the encapsulation of systems, embedding models, re-rankers, databases, and LLMs that allows AI agents to learn and adapt with new information over time rather than starting from scratch every session. How there are four types of agent memory drawn from human cognition, episodic, semantic, procedural, and working memory. He talked about how he advocates for a memory-first agent harness, and a key principle he emphasizes is reducing cognitive load for both LLMs and developers, for example, by consolidating into a single database that handles vectors, graphs, relational, and spatial data to avoid the anti-pattern of gluing together multiple databases.

1:00:51Jon Krohn:And he made a prediction that the flattening of AI engineering roles will happen in the near future where the future developer will need end-to-end understanding of the full agent stack. All right, as always, you can get all the show notes, including the transcript for this episode, the video recording, any materials mentioned on the show, the URLs for Richmond's social media profiles, as well as my own at superdatascience.com slash 985. Thanks to everyone on the Super Data Science podcast team, our podcast manager, Sonja Brejevic, media editor, Mario Pombo, partnerships manager, Natalie Zajski, researcher, Serge Massis, writer, Dr.

1:01:27Jon Krohn:Zahar Karchet, and our founder, Kirill Aromenko. Thanks to all of them for producing another excellent episode for us today for enabling that super team to create this free podcast for you. We're deeply grateful to our sponsors. You can support the show if you're listening to this by checking out our sponsors links, which are in the show notes. And if you yourself would ever like to sponsor an episode, you can find out how at johnkrone.com slash podcast. Otherwise, please help us out by sharing this episode with folks that would love to hear about agent memory, review the episode on your favorite podcasting app or on YouTube, subscribe if you're not already a subscriber, but most importantly, just keep on tuning in.

1:02:07Jon Krohn:I'm so grateful to have you listening and I hope I can continue to make episodes you'd love for years and years to come. Till next time, keep on rocking it out there and I'm looking forward to enjoying another round of the Super Data Science Podcast with you very soon. Thank you.