DevReal: Vasilije Markovic Interview

hello everybody I'm Alexi kraber the hello everybody I'm Alexi kraber the founder and organizer by a AI the founder and organizer by a AI the longest running deepest technical a longest running deepest technical a metup in the world here on location at metup in the world here on location at Cloud flare we have a amazing agentic Cloud flare we have a amazing agentic metup and the latest addition is vasil metup and the latest addition is vasil marovich who's the founder and COO of marovich who's the founder and COO of cogne uh start up doing AI memory what cogne uh start up doing AI memory what is cogne cogne is a memory manager for is cogne cogne is a memory manager for AI apps and agents it lets you load AI apps and agents it lets you load actually your data from your business actually your data from your business productionize it and then let agents and productionize it and then let agents and AI apps access it and enrich their AI apps access it and enrich their context so they can give you the answers context so they can give you the answers that your business needs and you talk that your business needs and you talk about it as memory and almost like a about it as memory and almost like a part of cognition how did you come to part of cognition how did you come to this idea that you know AI agents need this idea that you know AI agents need memory yeah so my background's in memory yeah so my background's in business and Clinical Psychology so business and Clinical Psychology so effectively I've thinking been thinking effectively I've thinking been thinking for a while how can we create these for a while how can we create these types of cognitive processes inside of types of cognitive processes inside of the vector and graph stores so we can the vector and graph stores so we can enrich the data store the information in enrich the data store the information in a way that humans process the data and a way that humans process the data and then effectively have a way to retrieve then effectively have a way to retrieve that in the way also humans access their that in the way also humans access their different memory elements so it's strug different memory elements so it's strug is you know really missing piece because is you know really missing piece because previously like when you chart with some previously like when you chart with some kind of chart GPT right it almost feel kind of chart GPT right it almost feel fairly stateless right and like when you fairly stateless right and like when you drop it you probably lost your session drop it you probably lost your session right it's gone right it's like now it's right it's gone right it's like now it's like it get it's getting better because like it get it's getting better because you kind of remember what we were you kind of remember what we were talking about right but then it's lost talking about right but then it's lost so so uh would your system allow this so so uh would your system allow this you know agents to preserve the state of you know agents to preserve the state of the work like how long can it persist the work like how long can it persist how big can it be so it can get to how big can it be so it can get to millions and millions of documents it millions and millions of documents it can persist the data from any agents and can persist the data from any agents and how imagine it is that we have layers of how imagine it is that we have layers of memory so we effectively store the memory so we effectively store the traditional information from databases traditional information from databases then memory interactions then uh then memory interactions then uh preferences of users and all of these preferences of users and all of these layers get accessed by the agents layers get accessed by the agents dynamically and these layers get updated dynamically and these layers get updated and changing similar to human mind so and changing similar to human mind so how we effectively update our knowledge how we effectively update our knowledge of the world as we go through it very of the world as we go through it very good and you told me that you know you good and you told me that you know you want to go back to school and learn want to go back to school and learn cognitive science how does that kind of cognitive science how does that kind of inform your work why would you study and inform your work why would you study and where would you focus your study yeah where would you focus your study yeah that's a good question so I already went that's a good question so I already went back to study I'm now third year of of back to study I'm now third year of of the of the Bachelors again with the 20 the of the Bachelors again with the 20 year old that's very fun and uh year old that's very fun and uh effectively as I go back to study what effectively as I go back to study what happened and what I saw in Psychology is happened and what I saw in Psychology is that they borrowed a lot of Concepts that they borrowed a lot of Concepts from uh computer science in the 60s from uh computer science in the 60s thinking of humans as thinking machines thinking of humans as thinking machines but then because they hadn't didn't have but then because they hadn't didn't have like these analogies to explain things like these analogies to explain things further because thinking machines further because thinking machines stopped at a certain point but they need stopped at a certain point but they need to explain behaviors and actions there to explain behaviors and actions there is a lot of patterns and systems that is a lot of patterns and systems that they built on top of the thinking they built on top of the thinking machines which now we can use back in AI machines which now we can use back in AI an example of that would be how do we an example of that would be how do we actually uh you know um um become more actually uh you know um um become more risk averse as uh we move and become risk averse as uh we move and become older so there is a mod set of models older so there is a mod set of models you can actually add that show that as you can actually add that show that as you get to age of 60 you're going to you get to age of 60 you're going to prefer not to do any risky thing prefer not to do any risky thing compared to like when you're 30 and you compared to like when you're 30 and you kind of like to do things but then kind of like to do things but then effectively these things we can put back effectively these things we can put back into the AI now and we can model how into the AI now and we can model how humans do it so the reactions of the AI humans do it so the reactions of the AI systems are more towards what we expect systems are more towards what we expect you know what's super funny to me like you know what's super funny to me like you mentioned thinking machine several you mentioned thinking machine several times today Mira morati the former city times today Mira morati the former city of and I announced a new company called of and I announced a new company called thinking machines and then what really thinking machines and then what really struck me most people don't even realize struck me most people don't even realize that there was an iconic company called that there was an iconic company called thinking machines in the '90s which thinking machines in the '90s which built a connection machine right so this built a connection machine right so this is like a recurrent thing it's probably is like a recurrent thing it's probably our memory is very short like when need our memory is very short like when need to put our member in cogy then we to put our member in cogy then we probably would remember right that probably would remember right that thinking machines is a recurrent topic thinking machines is a recurrent topic and people want to make this machines I and people want to make this machines I think this has been tried before and if think this has been tried before and if we talk about Brin it's very ambitious we talk about Brin it's very ambitious right it's a bit too much but what we right it's a bit too much but what we are trying to do at that this point is are trying to do at that this point is to create a better context for llms to to create a better context for llms to give it like as much data as we can to give it like as much data as we can to avoid the context issues and to make avoid the context issues and to make that auto updating as much as we can I that auto updating as much as we can I think thinking machines are like 5 10 think thinking machines are like 5 10 years ahead Mira morati is like from a years ahead Mira morati is like from a from a town opposite to where my family from a town opposite to where my family is from so neighbor in a way right like is from so neighbor in a way right like I think she's a bit more advanced I think she's a bit more advanced neighbor than than me at this point but neighbor than than me at this point but let's see how that that's right the let's see how that that's right the former venan State produced a lot of former venan State produced a lot of different cultures next to it exactly different cultures next to it exactly right so uh and uh you're at the state right so uh and uh you're at the state where you have this really robust uh where you have this really robust uh model and it helps connecting different model and it helps connecting different systems so you're basically looking for systems so you're basically looking for partners design Partners who will use partners design Partners who will use cogne right to validate the ideas so if cogne right to validate the ideas so if anybody here would like to try this out anybody here would like to try this out what should they do yeah they should what should they do yeah they should just go to our website cy. and check our just go to our website cy. and check our GitHub repo it's all open source reach GitHub repo it's all open source reach out to us on Discord send me an email out to us on Discord send me an email add me on LinkedIn happy to chat happy add me on LinkedIn happy to chat happy to try cogni different setups that at to try cogni different setups that at Value already right like basically yeah Value already right like basically yeah it's already better than a basic Rag by it's already better than a basic Rag by 37% we have the numbers benchmarks so 37% we have the numbers benchmarks so happy to share that with others happy to share that with others fantastic looking forward to your fantastic looking forward to your lighting talk and welcome to the middle lighting talk and welcome to the middle thank you very much