Podcast: Play in new window | Download
Subscribe: Apple Podcasts | Spotify | Amazon Music | Android | Youtube Music | RSS
Your organization’s brand is what people say about you after you’ve left the room. It’s the memories you create that determine how people think about you later.
Andrea Volpini says that the same dynamic applies in marketing to AI systems. Modern brand managers, he argues, need to understand how both human and machine memory work and then use that knowledge to create digital memories that align with how AI systems understand the world.
We talked about:
- his work as CEO at WordLift, a company that builds knowledge graphs to help companies automate SEO and other marketing activities
- a recent experiment he did during a talk at an AI conference that illustrates the ability of applications like Grok and ChatGPT to build and share information in real time
- the role of memory in marketing to current AI architectures
- his discovery of how the agentic approach he was taking to automating marketing tasks was actually creating valuable context for AI systems
- the mechanisms of memory in AI systems and an analogy to human short- and long-term memory
- the similarities he sees in how the human neocortex forms memories and how the knowledge about memory is represented in AI systems
- his practice of representing entities as both triples and vectors in his knowledge graph
- how he leverages his understanding of the differences in AI models in his work
- the different types of memory frameworks to account for in both the consumption and creation of AI systems: semantic, episodic, and procedural
- his new way of thinking about marketing: as a memory-creation process
- the shift in focus that he thinks marketers need to make, “creating good memories for AI in order to protect their brand values”
Andrea’s bio
Andrea Volpini is the CEO of WordLift and co-founder of Insideout10. With 25 years of experience in semantic web technologies, SEO, and artificial intelligence, he specializes in marketing strategies. He is a regular speaker at international conferences, including SXSW, TNW Conference, BrightonSEO, The Knowledge Graph Conference, G50, Connected Data and AI Festival.
Andrea has contributed to industry publications, including the Web Almanac by HTTP Archive. In 2013, he co-founded RedLink GmbH, a commercial spin-off focused on semantic content enrichment, natural language processing, and information extraction.
Connect with Andrea online
Video
Here’s the video version of our conversation:
Podcast intro transcript
This is the Knowledge Graph Insights podcast, episode number 27. Some experts describe the marketing concept of branding as, What people say about you after you’ve left the room. It’s the memories they form of your company that define your brand. Andrea Volpini sees this same dynamic unfolding as companies turn their attention to AI. To build a memorable brand online, modern marketers need to understand how both human and machine memory work and then focus on creating memories that align with how AI systems understand the world.
Interview transcript
Larry:
Hi, everyone. Welcome to episode number 27 of the Knowledge Graph Insights podcast. I am really delighted today to welcome to the show Andrea Volpini. Andrea is the CEO and the founder at WordLift, a company based in Rome. Tell the folks a little bit more about WordLift and what you’re up to these days, Andrea.
Andrea:
Yep. So we build knowledge graphs and to help brands automate their SEO and marketing efforts using large language model and AI in general.
Larry:
Nice. Yeah, and you’re pretty good at this. You’ve been doing this a while and you had a recent success story, I think that shows, that really highlights some of your current interests in your current work. Tell me about your talk in Milan and the little demonstration you did with that.
Andrea:
Yeah, yeah, so it was last week at AI Festival, which is a very large event with I would say hundreds of speakers. And my talk was about memory as a new framework for marketing in the age of AI assistant. And so I did a small test with the audience and I imagine we had a crowd of maybe, I don’t know, 40, 60 people attending the talk and a few others online. And I had these slides where I challenged the audience to program the memory of Grok. Grok is X AI system. And I wanted to do this with Grok and ChatGPT by asking the audience to share feedback about my talks. The talks was ready towards the end. And so I asked, “Okay, just share openly on X and Facebook about how was this talk?” And then we set up a small poll on X to let people simply vote if it was good or bad or relevant or boring.
Andrea:
And so we created engagement over social and of course, particularly because I’m still one of the few left on X, we interacted on X. And then all of a sudden, maybe after just a few minutes, one of my colleague went on Grok and asked, “What are the best talks at AI festival 2025?” And you can imagine there are hundreds of speakers, but Grok responded, “One highlight is a CyberAndy presentation that talked about using memory with AI system, and one of the attendees described it as mesmerizing, suggesting that he explored neuroscience,” and blah, blah, blah. So I was able to get there and to build memory collectively by having user share feedback on social network. And by the way, the same applied to ChatGPT. So asking the same to ChatGPT would also highlighted my talk versus many others, better talks on that day.
Larry:
That’s really one of the common observations and criticisms of LLMs has been their inability to access real-time information. That you build the model and there it is. So there’s obviously something going on under there. You’re one of the first people I’ve talked to who talks a lot about memory in these architectures. I guess, maybe if you could, I mean there’s so much going on in the last couple of years with this, but what have been the evolutions in the AI and LLM sphere that kind of have led to the emerging importance of memory in these architectures?
Andrea:
So I mean, I think all of us are realizing with daily use that we’re not interacting with language models anymore, but we are interacting with more complex systems that take into account multiple pieces in order to provide an accurate response. And every system, whether we’re dealing with Perplexity, ChatGPT, or Gemini, or Grok has its own different way of combining information in order to respond to us.
Andrea:
And so I started, because my work in marketing, I started to think how we should approach a customer that is becoming an AI. And then that was my trigger was like, okay, what if the next customer is not a human? What happens? And the first consideration to be made is that in the context of SEO, for example, we transition with after a few years from the idea of keywords and focusing on what are the keywords that I should rank for to focusing on the search intent of the user that makes a request to a search engine. But then all of this is gone, if I have to deal with ChatGPT, Deep Search, all of these disappear if I have to deal with something like Operator or Gemini Deep Research functionality because in the end there’s not going to be a human that it’s making the request, but it’s going to be an agent. And so I started to think, okay, what is marketing then if keywords are gone and also search intents is gone, what is left? What influenced the systems? And then I got to the revelation of memory.
Larry:
Okay. That’s really interesting. The way, that evolution you just described too. The one thing that occurred to me as you were talking about that is that ostensibly Google has always favored that if you’re doing things that appeal to human beings, you’ll rank better in the search engines. But it sounds like from what you’re saying, and so that kind of guided SEO for the last, I don’t know, 15, 20 years, but now you’re saying we’re in this, we’ve kind of switched to where, and so I think a lot of SEOs, the perception was they were just playing to Google to trying to game Google’s algorithms. And it’s not like gaming, but it’s understanding your audience. It’s like any old communication problem, understanding your audience.
Larry:
So what are you seeing as the difference as you make that leap from search intent to memory needs of these new like Deep Research and tools like that? How do you, and your end goal in this is to automate marketing tasks. What does that look like? What’s the pipelines or procedures or your approach to that?
Andrea:
So I started from building our system for our client to let’s say improve the quality of content recommendation on an e-commerce website or increasing the quality of internal links and doing that at scale required an agentic approach. So there is a language model driven agent that has to find relevant pages and then has to have the notion of what is a main query for these pages, and then as to learn how to craft a proper anchor text in order to link one page to a relevant other page.
Andrea:
So as I was doing this development, I realized that the essence wasn’t really the model itself. That of course has its own characteristics and biases, but it was really the context that I was feeding the model with in real time in order for it to do the task. And so I realized that a pivotal change, it’s on how we craft these memories. What is the information in context that we want to pass to the agent in order to do the task properly, and how does the system evolve as things move forward and user maybe start clicking on these links and search engines start crawling these pages. And so I realized that memory was really the underlying element of success for my AI agents.
Larry:
So memory, and when we think of memory, you think of RAM and the computer memory, but also human memory and the different kinds of memory like short-term, long-term. And then you talk about there’s three kinds of main buckets of kinds of memories. Can you talk a little bit about that and kind of start to draw the parallels between human memory and how this is manifesting in these AI architectures?
Andrea:
Yeah, so of course we as human have short-term memory and long-term memory, and the same applies to an AI system. Short-term memory, it’s traditionally represented by the so-called context window of a language model. And of course different system have different context window. If I choose to build an AI agent with a model like Gemini, I would leverage on a 2 million tokens context window. And if I manage to use, for instance Grok, I will have 1 million tokens, whether with ChatGPT, the context window shrinks and gets down to 128,000 tokens. So the short-term memory, it translates to in context for language models. Now the long-term memory can be a knowledge base, in our case, a knowledge graph or search, a web index of information that it’s published on the web and it’s crawled and indexed with a modern search engine.
Andrea:
Now when starting to do this exercise, when I realized that in order to make my AI smarter, it wasn’t really about the best model, but it was about the context. I started to study how the brain works and of course the brain as we can simplify things and talk about old brain and new brain or system one and system two. And I focus on memory that it’s a function of the neocortex. And I found a very beautiful framework in a book titled 1,000 Brain Theory by Jeff Hawkins.
Andrea:
And in these conceptual framework of how human memory works, I found a lot of similarities with the work that we were doing with our marketing ontologies and SEO ontology. Because the brain in order to, for example, describe a glass of water, would create memory by using cortical columns. Imagine these are spaghetti into our neocortex and each one of these columns, it’s creating a representation of the world of the glass of water in that case. And of course there is the tactile feeling. There is what I’m perceiving with my sight, the smell, the context in which I’m holding my glass of water and each column independently it’s creating a knowledge representation and the different knowledge representation then converge by voting into a unified understanding of the glass. And that would create the memory and it’s living because it’s working by predicting what’s going to happen next. If I turn it on the right, maybe the water will go down, and if I move it close to the mouth, maybe I just want a drink.
Andrea:
And so memory, it’s essential in the way in which we as human learn and adapt to the environment. And therefore it should have been, it is becoming quite significant elements of current AI system. And so as a designer of AI system, I realized that memory was the thing. It wasn’t really about just training the model, but it was really about feeding it with the right memory and the right memory of course, it’s encapsulated into the pre-training. For example, the model learns during pre-training about all sorts of things. The entire web becomes part of its memory, but then the model keeps learning when it interacts with us in context. And so the way in which the model differentiates itself, it’s really by the way in which memory are built and consumed by these systems.
Larry:
As you’re talking about this, I’m picturing these memories just floating around, but they’re, like you said, they’re a form of knowledge representation, and also a lot of jumbled thoughts in my head as I think about this when you talk about those cortical columns, each of those is an attribute or a facet of the thing you’re talking about and together they give you this impression of what a glass is.
Larry:
And that’s a really fascinating look at how the human brain works and I’m kind of inferring from what you say, how that manifests in these architectures, but how does that look in terms of, like you said, you talked about the context and then how the memory and those two kind of working together, very much like Kahneman systems one and two kind of thinking how does it come together with the architecture? I’m still trying to get my head around how you stitch together these bits of knowledge into, and both knowledge and then the systems one side of it, the vectorized proximity things that are floating around as well. How does this come together to do-
Andrea:
In layman’s term, AI, it’s about knowledge representation. Solving the AI challenge to build intelligent machines, it’s a knowledge representation problem. Now, one thing that we learned so far, even in our direct experience is that a given entity, an object, much like the multifaceted representation in our cortical columns, can be represented with discrete fact, how much water do I have in the glass and what material the glass build on?
Andrea:
But then there is something that it’s more nuanced. What is my feeling when I approach a glass of water in that specific point in time? And so these can be represented a continuum, in an embedding form. And so right now in our knowledge graph, we represent entities with both triples and vectors. And this is very similar to how of course, for example, Perplexity would represent knowledge. Perplexity shines on finding up-to-date information using its vector-based search. And so it’s primarily focused on getting real-time updates. But of course, if I want to buy a pair of shoes, I would rarely rely on Perplexity because it doesn’t have in its infrastructure and architecture a shopping graph like the grounded Gemini 2.0 Flash S. And of course a giant shopping graph what Google has built over the years gives to its AI, the advantage of having a very specific memory for tackling transactional intents.
Larry:
That gets into also, again, for a marketer or anybody consuming this kind of knowledge on the web, understanding the capabilities and limitations of any one model. I’ve already played a fair amount with the tooling that lets you pick which model you’re going to use. And I’ve just been doing a kind of willy-nilly to experiment with whatever people are talking about in the media. But now I’m thinking that we need a knowledge graph with all the information about all the various models so that you can pick, if you’re shopping, you would never use this model, but you’d always use that one.
Andrea:
Everything that is transactional and entity based requires a specific understanding and amount of high quality data that Google has, for example, for shopping items that Amazon has. So we can expect to have model from Amazon excel at transactional intents. We don’t go right now on ChatGPT to ask where to get the coffee the next day simply because the local entity database that currently ChatGPT has, it’s very limited and God knows where it’s coming from. It’s a combination of what Yelp and whatever it’s grasping in real time from Google, or Bing but it is not reliable enough for us to count on it.
Andrea:
But for example, ChatGPT is very personal in the relationship. ChatGPT has been first with Gemini to build memory from one conversation to another. And so it can recall what we have discussed in the past. You can ask to ChatGPT to draw a picture of you, and based on all the conversation that it has, it would create a beautiful profile that would be most likely accurate based on all the conversation that he had with you. So it’s very personal, and Perplexity has no personality, has no personal contact with you. It only works extremely well with its embedding vectors and ability to trigger search for giving you the latest update. So the memory, it’s different.
Andrea:
A system like Grok is a giant context like Grok has 1 million token. So within the same conversation it as it will have a great recall of what you have said in the different trends, but the strength of Grok is its access to the live social media data of X. Therefore, if I want to build the memory for Grok, as long as I have users engaging on Grok, that’s was my Milan experiment. What is the memory that Grok uses the most? Its social media access to X. And so if I create engagement on X, the model in real time will react.
Larry:
Yeah. And as you’re talking there, something you just said triggered this memory, and I want to make sure we get to this, is that in addition to that kind of highest level distinction between short-term and long-term memory, there’s kinds of memories and from what you’re saying and those distinctions between procedural and semantic and wait, what’s the third one? But the different kinds of memory, what are the implications for those in these architectures? Like the, oh, the temporal one to the, tell me about that.
Andrea:
Episodic.
Larry:
Episodic. Thank you. Yeah.
Andrea:
So AI agents rely on different memory frameworks. So for me, when building an AI that automates SEO, my memory is built on this ontology that we call the SEO ontology. So this is the blueprint of how we envision memory for an AI agent that does SEO. So depending on the task, an agent will need to access to different memory types. So one memory type is the semantic type of memory. So the memory that, for example, Gemini will need when I asked to find the best pair of running shoes.
Andrea:
A semantic memory can be organized in a graph in the form of entities and of course the retrieval of a semantic memory can be done with a combination of vectors and attributes and triples. So we can much likely access semantic memory by combining the triples in the graph with the vectors.
Andrea:
The episodic memory, it’s the things that the system has learned, whether it has learned things about you or it has learned things because there has been an evolution over time. We could represent an episodic memory with an event ontology. On that day, Andrea spoke about his love for mountains. On the other day, Andrea mentioned his wife. So time-based memory help the agent understand that there is a flow on time.
Andrea:
Imagine that the limitation of a language model is that it’s designed to be stateless. So after pre-training, everything that has been learned, it’s encoded into these billions of parameters that cannot be adapted dynamically. So in order to make the system dynamic, we have to feed it with memory. So episodic memory, it’s for example, the memory that in our case when we build our agent, would allow the agent to remember that that workflow was successful with that type of site. So with the episodic memory, in our world of marketing automation, the system learns what worked and when, and so it can learn. And of course, imagine that these systems are lively updated. So if data comes from the Google Search Console about people requesting a page, using a query, I immediately bring this into the knowledge graph and feed it into the agent. So memory is dynamic and helps the system to adapt and evolve, much like the human brain.
Andrea:
The other memory, it’s procedural. How do I do certain things? How do I make pasta, how do I build internal links? There are sequences, there are steps, and if I have to represent it in a graph, I would possibly rely on a class like a schema how to where I can describe the steps that needs to be taken for achieving a specific goal.
Andrea:
So semantic, episodic and procedural memory are the three type of memories that this system use. This is important for us when we build the system, but it’s also important for us to understand when we interact with something like ChatGPT, we’re not interacting with the language model, we’re interacting with a complex architecture that has memory, and of course ChatGPT excels with episodic memory as we mentioned because it can keep track of the conversation, but for example, lacks semantic memory when it comes down to products or local businesses.
Larry:
As you say that, I’m really curious now about even among advanced practitioners in AI architecture, construction, a lot of this will be new or different from the way they’ve been doing things. Has memory been considered in architectures now, or is this truly a new thing that folks are going to have to adapt to and work into their products?
Andrea:
So in the world of autonomous agents, it’s becoming more common to think about what is the memory framework that you’re going to use and how a knowledge graph can help. Of course, this was very close to my area of expertise because we work in graph technologies. And as we started to add the vectors into the knowledge graph, I realized that it was just the perfect infrastructure for creating a memory layer for AI. But then I also look on the other side, I look at, “Okay, how can my client be more visible on conversation? How can I help the brands using our solution to be more present when interacting with the large language model or an AI system?” So I started to think, okay, well, we can’t work on search intent only because that’s for the humans, but we’re directly interacting with agents. And so based on my experience, I started to think of marketing as a memory creation process.
Larry:
That’s really, and so this sounds key because theoretically 2025 is going to be the year of the agent and agentic architectures and all these things. Do you see this new memory capability as key to that?
Andrea:
Oh, yeah. Oh, yeah. I mean, there’s no way to build an intelligent system other than by providing the model with an understanding of the world, a way of reasoning about the task and access to memory, whether in the form of a web index, in the form of a knowledge graph, in the form of a knowledge base or a combination of all three things. And so the idea of an external knowledge base that was introduced with the retrieval augmented generation to solve the problem of the knowledge cutoff of language model has evolved in a way richer scenario where these nonparametric memory are becoming strategic for the success of that specific agent.
Larry:
This seems really, well, we should check back in six months and see how this is all panning out. And I assume your whole workflow will have grown even more by then. But hey Andrea, I can’t believe that we’re coming up close to time already. But before we wrap up, is there anything last, anything you’d like to revisit from the conversation or just make sure we share before we wrap up?
Andrea:
I think that what is important is that I want people to start focusing on the importance of creating memories for their brands. And I want people to start looking at understanding deeply how a language model perceive their brands and product and services. Because as we move forward, I don’t expect to see static content anymore. I expect to see an hyper-personalized level of conversation where in the end, memory is going to last much like the tone of voice of the brand is going to last. And so I really want people to start focusing on creating good memories for AI in order to protect their brand values.
Larry:
Oh boy, that’s so interesting. There’s a whole other conversation we need to have with Jason Barnard about branding and stuff. Well, very, very cool. Oh, one very last thing, Andrea, if folks want to follow you or connect online, what’s the best place to find you?
Andrea:
Yeah, I’m on LinkedIn, I’m there as Andrea Volpini. We do have also quite a rich newsletter from WordLift on LinkedIn and X. I’m CyberAndy. It’s a little bit shameful for me to be there, but I am still there. I’m also on BlueSky, I think. But yeah, I don’t spend too much time there, to be honest, but yeah.
Larry:
Yeah. But I hope you grabbed the CyberAndy handle at each of those.
Andrea:
Yes, I did. I did. I’m there. I’m there. I’m there as CyberAndy.
Larry:
Excellent. Well, thanks so much, Andrea. It’s always great to talk and this was a particularly fun one.
Andrea:
That was good. Thanks. Thanks, Larry.
