Full Transcript

Dhruv Sharma: Hey there listeners, how is Friday coming along? Are you guys watching the World Cup? America is celebrating its 250th birthday tomorrow, by the way, and here we are streaming number 108 and chatting with Jay and or Jessica, who are building two products at once, Shrum and Minimi. Guys, welcome to the show and we'll have you, we'll have to maybe just start with a story that we were chatting about just a few minutes ago. Tell us, you know, who came first, Shrum or Minimi?

Jay Gadekar - Co-founder & CEO, Shram: Sure, so Shrum came first. Shrum has been the product that we have been building for a while now. Minimi has been a recent experiment that just went really very well.

Dhruv Sharma: Tell us more, you guys have like this crazy way in which you've been acquiring users. Tell us the whole story and then we'll get started.

Jay Gadekar - Co-founder & CEO, Shram: For sure. So about one and a half months ago, we heard from a lot of our users, okay, you guys have built this ambient capture and on-device memory layer that you're using inside Shrum to, let's say, find tasks for me, which is what the previous Shrum used to do. They requested repeatedly, why don't you guys give me access where I can just chat with my memory? It's like your second brain. So this was also the time when we lost our pet kitten. So that gave us a few days of headspace, okay, let's just go and try building the site project. And then we wrote an article about how your memory is not a product but a feature. That article had 50,000 readers, some of whom were from XAI, Google Emergent Anthropic, and they wanted to try the product. So we prioritized it, we made it into a nice Mac app. The way it works is it has context of everything that you have ever read, heard or said on your MacBook, and those memories get stored right on your device and you can plug it into any LLM, let's say Cloud. So the next time you go to Cloud, you don't need to copy paste anything, you can just chat with your memory without uploading anything. And that's basically the product.

Ojasvika Sahu - Co-founder & Chief Design Officer, Shramm: Just to give you an example, what you can basically end up doing is, now that Malini has access to every sort of thing that you've been doing on your screen, and it's all going to store locally, it's not going to do screen recording of any kind, so it's very privacy first. What it does is it essentially captures all of that, puts it in a small bottle, that's called Malini, that's our memory layer, and you can connect it to Cloud and essentially ask it anything like, what were the three things that I missed doing yesterday? And Cloud will be able to answer that. It won't be able to answer that without Malini.

Jay Gadekar - Co-founder & CEO, Shram: Yeah, and then we spoke about some of these use cases on X through Odhisheka's Twitter account, and we went viral a couple of times. So that brought us users from all over the world. They started converting real fast, so we just thought, okay, why not for the time being focus on what's happening right in front of us? And that's where we are at right now. So while we do have two products, one of them is what's getting most of our attention right now, which is Malini.

Utsav Somani: What were the use cases that went viral?

Ojasvika Sahu - Co-founder & Chief Design Officer, Shramm: Oh yeah, so that was actually an interesting story. The first one that went viral is, I literally started, it was an overnight experience, you know, Fable 5 was live for a couple of days, and I just wanted to know what's the maxed out version of Fable 5 plus, let's say what I can do with Malini. And Granola is one of my favorite products, I absolutely love that product, it's so essential and so simple. But I just thought that now that Malini has access to all the calls that I'm anyway doing, might as well just build my own Granola. And that's the first one that completely blew up, where I just said that now you can basically kill Granola and you don't have to pay for a separate subscription anymore. So that was one. The second one was even crazier, because that went even more viral. I think Satya Nadella was sort of talking about how, you know, the frontier tech is amazing, but serenity is very critical for people, you need to own your own data. So I felt that because your memory is with Malini's on your Mac, technically it's sovereign because it's your own data, and you're putting it in Claude. So your output is as good as the data it has. So we thought what if you could simply connect my Malini to let's say Jay's Malini or any other teammate of mine and their Malini. And I just give Claude those three brains and I can essentially ask it what have these guys been working on and it will be able to answer that. So that was something that you know, people absolutely love. So it's a mini-me is like a piece of clay, like sort of a toy that you can do anything with. And I think people are really enjoying that.

Utsav Somani: I mean, you mentioned Granola, that's very interesting, because OS level AI products are fairly exciting, like, but what enables them? What's the technical architecture? How is it? I mean, how is this possible now?

Jay Gadekar - Co-founder & CEO, Shram: Sure. So Granola taps into your system audio and your microphone. These are two permissions you can give it on your let's say Mac. So earlier, right, like with let's say Fireflies, you would have a board join your meeting. And there was one take that people didn't like too much. And Granola immediately vaporized that entirely because it's tapping into your audio.

Utsav Somani: What do you need to happen? Because I mean, there must be a lot of processing that goes into like, I mean, just reading and just consuming everything that's happening on the screen, right?

Jay Gadekar - Co-founder & CEO, Shram: Okay, that way. Yeah. So sure. So when it comes to listening into meetings, that's what like Granola and Minimi both do, which is tapping into these two permissions. There's a third thing, right? Like when we started Shrum, we were experimenting a lot with screenshotting. So every few seconds, we would take a screenshot. But there are three challenges with that. The first one is the frequency issue, because you can only do it for, like, let's say every three seconds. So there's a blind spot between those two moments. The second is that it's very expensive and time consuming for any LLM to process it. And the third one is that it costs a lot of money. But we use actually accessibility infrastructure, which is built into your MacBooks. So it's all instantaneous, free, and extremely accurate, like it returns data, almost integration grade. So that's the infrastructure for the ambient capture. Then we have the memory layer, which is basically a vector space on your system, it is a vector database, everything that accessibility or your audio returns is chunked into small embeddings, and they are all stored in this vector space. So every time you try to find, let's say, when was the last time I spoke about a cat, for example, it will go and try to find the word cat, then it will bounce it off and maybe find more related concepts like ball, or I don't know, litter, and so on. And that's how it starts retrieving information from these embeddings. So a large article gets chunked into smaller embeddings, and then we retrieve one embedding and that embedding leads us to the main article. And that's basically the architecture behind the memory layer. So these two things together, basically what Minimi is.

Dhruv Sharma: I want to get your thoughts on a few different things, right, because we've already covered a lot of ground. The first thing is, you know, in the world of AI, everything changes so rapidly and changes all the time. The consensus keeps shifting, you know, today, you'll hear that prompt engineering is what everyone needs to good at, then it goes to harnesses, and then context engineering and loop, the terms keep changing, the concepts keep changing, everything is always changing, the consensus keeps shifting. I think with respect to what you guys are doing, like this, this idea of a second brain, we've had this idea for quite some time now, remember, there was a lot of excitement around a product, you guys must have also tried it called Rome research, right? So one, I want to get your thoughts on how do you keep up with all this shape shifting that keeps happening in the in the in the AI domain, right? How do you discard your previous beliefs? How do you keep updating your priors? The second is, you know, what is your version of an ideal second brain for individuals? And what's the equivalent for organizations? Some people call it a context graph. Is that what you guys think is what needs to exist inside of an organization? Yeah, and then we'll, we'll get to more topics as well.

Jay Gadekar - Co-founder & CEO, Shram: For sure. So firstly, how do we keep up? So I would say whenever a new tech comes out, generally, it's too early. That also happened with second brains, sort of that movement that happened about a year ago, everybody came to build a second brain. And technically, they were not wrong. It just so happens that if you look at a cloud today, or any other LLM provider, these folks are spending a lot of their energy on building persistent memory. So every time you have a conversation with your LLM, they remember that conversation, what we are banking on rather is everything that is not part of that conversation. So a good example is, let's say you integrate your email, or I don't know your linear into your cloud. It has context to all of that. But what happens about your conversations that are happening on your LinkedIn, or your x or your WhatsApp, they don't have any direct integrations. So Claude doesn't know that you sort of made a poster on Figma, there is no way for it to know. But when you do something like mini me, it's ambient capture.

Dhruv Sharma: So you're saying like, Claude only knows what Claude knows, mini me is equivalent of you know, this parrot that sits on my shoulder and observes everything I'm doing on my device. And that's what we call ambient capture. And then it doesn't unlike the model companies and their products, it doesn't compress it into a memory file. It actually takes the context as is and creates local files without any distortion.

Jay Gadekar - Co-founder & CEO, Shram: Yeah, you're absolutely right. You can think of it this way that all LLMs have blind spots, no matter how many integrations you use, you plug in, these blind spots can only be cleared if you have ambient memory, which is what we have built. Now, the second part of what you said, yes, we store it on your device, it does not belong to any LLM provider. So going back to what we were just talking about server netting, right? You own your own memories. So you can plug them into any LLM you like, you are no longer linked, like you're not associated with a certain LLM. So let's say you have been using Claude for two years, it doesn't mean that you have to stay there. You can plug out your memory from Claude and put it into anything that you like. So tomorrow, as MacBooks are getting better tomorrow, when you have better local LLMs on your own device, the kind of architecture that we have built, you can just plug it into your MacBook. You don't even need internet and you have an absolutely fantastic second brain on your system. Is this also a second brain like the ones we saw a year ago? Yes. It's just that this one does not have any blind spots. So I would say the first question, how do we keep up? I guess we follow along all companies in this space very, very closely. Right now, we know there are only probably two or three companies in the world doing ambient memory capture. And we are very proud to say that we are one of the best out there. What we saw as a gap was that people don't like getting used to another separate app where their memory lives. They want to plug it where they live. And most of Minimi's users are already on Claude. So the interface is Claude and Minimi is basically ambient, it's invisible. So we picked that up and clearly people have been enjoying that. So I would say the way we keep up is by analyzing what other companies are doing and also very closely listening to our users. And the second thing, like you asked, what is the version of the company brain? I would say, Ojaswika's post about the company brain went viral, primarily because the principles on top of which...

Utsav Somani: You should get Ojaswika to write Twitter threads for us at TUM. She's gone viral a few times, it seems.

Ojasvika Sahu - Co-founder & Chief Design Officer, Shramm: It's something I'm figuring out why I haven't quite found the answer. But I think it's more about, I think people are loving the idea that something as simple as this could be a very powerful tool, because you can just like Jai mentioned, you can just plug and play it into things. So that's something really very interesting for most people.

Jay Gadekar - Co-founder & CEO, Shram: But yeah, I think Jai, you were saying something. No, nothing. I was just saying that the principle that you own your own memory and you can plug it anywhere. So even if tomorrow your company requires a company brain, you on Minimi are basically a node or a neuron of that brain. And that memory stays on your device at any given moment. It's something that people, it seems, appreciate it a lot. We think that's the right future for any tech like this, because with the right guardrails, filtering out any personal context, this becomes the ideal node that strengthens the company brain that could just be your Claude or your Gemini or your OpenAI in the future. So that's the tech that we are very excited to build. That's our version of the second brain or the company brain that we imagine.

Utsav Somani: What are some of the technical challenges that you're looking forward to solving in the next six to 12 months? Like this is OS level stuff, right? So, I mean, the ground keeps on shifting, Apple's building Apple intelligence and trying to build context within your devices as well across the iOS ecosystem. So how do you, I mean, deal with this on different operating systems and how do you think about what challenges do you want to solve next?

Jay Gadekar - Co-founder & CEO, Shram: For sure, I think the biggest challenge that we have had has been retrieving the right information from the memory layer. Frankly, right now, our benchmarks are 50% higher than the state of the art on the Beam evaluation test. So we are quite satisfied. So I wouldn't state that there are any challenges. We actually already have overcome the biggest hurdles so far. The next steps, although, are like you said, like, what if we have to do the same thing on not Mac, but Windows, right? How does that look like? That's currently a black box to us because we haven't really checked that possibility yet. The second thing I would say is we really want this to be completely on- device, meaning there's one API call that we made when we are building a memory that goes to a paid server we have on with, let's say, Gemini. We also want to eliminate that completely. Now, in order to do that, we need infrastructure around us, meaning that MacBooks need to get better, which they are, like the recent WWDC, they announced that there will be better local LLMs on your device. So we really want to sort of be the flag bearers of making memory completely on-device. And maybe in the near future, we might even open source our tech. So making it completely local remains to be our biggest, I would say, exciting point, I would say. Like, that's what you want to look up to. But no major speed breakers, I would say.

Utsav Somani: And how many people are using your product right now?

Jay Gadekar - Co-founder & CEO, Shram: So about a thousand people are already there. And they come from more than 20 different countries. And we are just talking to them, we have been talking to them recently, we are trying to get where should this head in the near future? What are you using it for? Many of them are hooked to it.

Ojasvika Sahu - Co-founder & Chief Design Officer, Shramm: In fact, one of the recent conversations we had was with Mr. Bill, he is an ex-NASA engineer. And I think he came out of retirement to sort of, you know, he's 73, he came out of retirement to just because he got excited about AI and he absolutely loves Minimi. He keeps talking about Minimi every day on Twitter. And I think that's the kind of...

Utsav Somani: And I mean, it's such a fascinating product, like, I mean, I would, I mean, after the show, maybe try it out for sure. But why are just thousand people using this? Why is it not like 50,000, 100,000 people right now?

Jay Gadekar - Co-founder & CEO, Shram: I would say we are over the loop. Yeah, thousand is the cohort that we are looking after. There are probably more users there. So even with Shrum, we are very specific about the kind of users we should be looking after. And what is it that they want? Otherwise, we're just bringing in unnecessary noise. So there are more people than a thousand. But we are looking very closely to them because they form a certain ideal customer persona that matters to us.

Dhruv Sharma: And so when you run this beta with these thousand power users, have they consigned 100% of their memory to Minimi or are they still using traces of, I don't know, Obsidian here or there or some other tool?

Jay Gadekar - Co-founder & CEO, Shram: That's a good question because we do have a friend of ours. In fact, he was using Obsidian and one day he just came by and told us, OK, your admin memory can really be helpful to my Obsidian. So the thing is about Obsidian, you have to maintain that manually. And with Minimi, you have the same outcome without the need to maintain anything manually. So is he still on Obsidian? Yes. But is he slowly moving more towards Minimi? Yes, that's true also. So maybe there's a degree of transition that's happening. But yeah, like we do have another user. He's called Charles. He's also like 50 plus. Now for him, it's been completely Minimi. He was on a certain different company's product that built memory for him. He realized and he ran his own benchmarks and he realized that we were four times faster in retrieval. So he just moved quickly on to Minimi and he has built some really complicated dashboards with it. So we are seeing some habit changing from where people had to manually create their memories to now completely relying on an automatic admin memory layer like Minimi. So we're seeing that happen for sure.

Utsav Somani: And I mean, what are some of the feedback that's coming in on like, I mean, retrieval of memory, like latency and like other technical stuff? Has there been any cases of hallucinations in memory also?

Jay Gadekar - Co-founder & CEO, Shram: Yeah, like there would be. If you're using Minimi, let's say 100 times, I'm sure a couple of times this will happen. Have people complained about it? Not once. We haven't really seen any, I would say, serious complaints about this. When it comes to the speed at which Minimi works, we are actually really very fast. So again, like I know like many in India, the way we think is, okay, SF has to lead the way first and then we implement something, a version of that right here in India. Frankly, in this case, we are leading that way. We know that we are the fastest out there. We know that our capture is really good. We know our accuracy is very high. So technically, actually, we are very, very sound. It's just about people's perception that, okay, this is not happening in SF, so something is possibly amiss. That's the only thing, yeah.

Dhruv Sharma: And more power to you as you do that. All right, let's ask the following question. As an end user, how will having always on an on-device memory change the quality of my life six months from now, a year from now? Like what possibilities does it unlock?

Jay Gadekar - Co-founder & CEO, Shram: Surely. So let's imagine a few months down the line, and since we are only on MacBooks, and that's also because they're really fast and innovating, let's say there's a very good...

Dhruv Sharma: So I'm sorry, but when I was chatting with someone on our team just before you guys joined, they're also very expensive and increasing hard to get, like, that's going to be an unexpected challenge you run into if you're only serving the Mac ecosystem. But you were saying, Jay?

Jay Gadekar - Co-founder & CEO, Shram: I agree. But yeah, I'll address your question first. So imagine like you have a very powerful local LLM on your Mac, and you don't have internet, and you just read an article. And if you remember, we actually joke about this. If you have seen that movie called Krish, right? And he's sitting in a room and he picks up a book and he just slips through it. And now he can answer, okay, on the 80th page, there was an error, right? That's the kind of stuff that Minimi can do for you. You just open a file, just go to your local LLM, ask it anything about it, it can do it for you.

Utsav Somani: The character Mike from Suits, basically, has that photograph.

Jay Gadekar - Co-founder & CEO, Shram: Yeah, yeah, yeah, sort of. And you can experience it right away today. Like it works. It's actually that good. Just go, I was reading an article, I went back, it knew someone I was chatting with, there was some context.

Utsav Somani: I'm going to try this out. How much do we have to pay for this?

Jay Gadekar - Co-founder & CEO, Shram: It's free. So you can go ahead and try. If you need more than five MCP calls a week, that's when you can try to subscribe to the higher tier. But most of our users, most of our non-customers, these folks are more than happy with just like five. So basically, once a day is what we provide for free for now. So we'll see. We are just still testing. We are trying to see what limit is great for a conversion wall to be the next thing. So you can go ahead and try it. And I'm happy to share some coupons with you guys so that you can try it for like for as much as you want.

Utsav Somani: Please do. I mean, we've got an active community of founders that we run, people who appeared on the show as well. If you like the product, we'll recommend it there also.

Jay Gadekar - Co-founder & CEO, Shram: For sure. We would love that.

Utsav Somani: Dhruv, any final closing question?

Dhruv Sharma: I'm going to have to think, Utsav, but if you have one, go ahead.

Utsav Somani: No, no. I had a good chat. I'm excited to try the product.

Dhruv Sharma: Excited to try it right away. Yes.

Utsav Somani: Dhruv, I was like, man, this is insane. I remember, I think day before or yesterday, I think Vaibhav was recommending from Better Capital in his WhatsApp group.

Jay Gadekar - Co-founder & CEO, Shram: Yeah.

Utsav Somani: Do you guys know him? Is he an investor?

Jay Gadekar - Co-founder & CEO, Shram: Yeah. I mean, we are in the same group as Better Capital. So we were there when he shared on that group. So yeah.

Utsav Somani: OK. And you're raised from Boundless as well?

Jay Gadekar - Co-founder & CEO, Shram: That's correct. Yeah. Boundless is our investor. That's correct.

Utsav Somani: Awesome. Perfect. Thank you so much. And wishing you all the best on this journey.

Jay Gadekar - Co-founder & CEO, Shram: Thank you so much. Thank you. Thank you, Dhruv. Thanks, Utsav. See you.

Utsav Somani: All right, listeners. That's it from us. Have a wonderful weekend. And we'll see you on Monday. And Monday, apparently, Spain and Portugal are playing each other as well in the World Cup. So I think it's going to be an exciting Monday night also after the stream. All right. Have a safe one. See you.