Building AI Agents with Memory with Dhravya Shah (Founder, Supermemory)

How long did it take to integrate into ComposeU?

We had a PR done in like 45 minutes and merged in an hour. Dumping will work, but it will make your agent worse in quality and worse at just answering the question in general. And this is one of the biggest problems with OpenCLAW, that sometimes you would just want it to invalidate some previous piece of information and have a new piece that's built on top of it. How we do it is we call them like, you know, the update edges. The agent will know that, okay, this was the prior one and this is the newer one, and the new one is built on top of it.

So we're here at Dhravya's place. He's the founder of Super Memory. We're gonna do an interview with him.

Hi, how's it going?

Hi, nice to meet you.

Yes, nice to meet you.

Thank you so much for inviting us. I'm really excited to chat. I love this space. Is this where you live and work?

Yeah, yeah, that's right.

That's so amazing.

Come in, let's sit.

So nice to meet you, Dhravya, finally. I love what you're building and I'm excited to hear more. Can you give a little bit more of an introduction of who you are?

Yeah, a little bit about me. I'm Dhravya, I'm the founder of SuperMemory. Um, have started SuperMemory about a year ago to solve like the, the memory problem of AI, and we can go more into that, but yeah, that's a little Amazing.

And how I found out about you— so I searched you up on LinkedIn for some research, and I saw that you were a part of BuildSpace Season 5.

Yeah.

Which is really exciting because I was also part of that. And I really want to know, did the early iterations of Super Memory start when you were at the 6-week program?

Yeah, yeah, so it did. So right before the 6 weeks of BuildSpace started, Season 5 was over the summer. I had like this, like this app I had made for myself and I had put it on GitHub. It was very new at the time. BuildSpace kind of literally guided me from making that into a, from project to something that people actually use, which is like a product and then into a company. So it was like an amazing experience for me. I've made so many friends during that time. It was in 2024. Super Memory became like one of the largest repositories or fastest growing repositories of 2024 because of BuildSpace and like all the people that are, that were helping each other.

So yeah. Wow. Wait, that's so awesome. How old were you when you, when you started this?

So I was 18 when the first iteration started. Okay. Now I'm 20, unfortunately. Oh my gosh.

So what inspired— is this something like personal in your life? What inspired you?

So it is actually both personal and Very, uh, very interesting. So I was, you know, like, the, the idea came from this fact that I had a note-taking app for myself which turned into a journal app. And I was, you know, like, writing a lot of letters to a lot of my friends on this letter-writing app. Very interesting. And I was like, dude, I'm writing like millions and millions of tokens of everything about myself. Yeah, what if I take this and turn it into a chatbot that is me. And so I made this thing for a friend's birthday, like just my own chatbot, which is kind of interesting. But, um, that was terrible at the time. And then I was also like, you know, I, I was not doing— I did not do high school, like I essentially dropped out. So I was like really knowledge hungry for anything I can learn. So I was saving all of these things to my note-taking app. Then again I was like, oh, maybe this should also be something that is about me. I should like, you know, this is my brain that I should be able to query. So I turned that into like— it was called Any Context before, which is all context about me. It took— then this random guy gifted me the domain supermemory.ai, which is also an insane story. And then that turned into Super Memory. So yeah.

Wow, what's the insane story?

Should we walk? Yeah, let's do it. So there was this guy at Meta, he was working on React Native, and he just saw my project in the wild like on GitHub and he then found me on Twitter where my latest tweet was like should I call this Super Memory at this point. Then he was like oh, I happen to own the Super Memory domain, I can give it to you for free if you let me invest in your first round. And I was like I was just a guy in college building this as a side project so I was like okay, obviously just give it to me. And then 2 years later when I did raise my first round, I let him know that I'm raising and he invested. Wow. So yeah.

That is so, so amazing. And so why do you think memory is such a big deal for AI agents? What breaks without it?

Yeah, so right now there's like 2 general problems here. So one is the context length being like either short or really bad when it's long. And the second is self-learning about the user and personalization is really hard. Um, so we, we saw both of these use cases. So on the context length front, there's this paper called NoLIMA, and then there's another called ContextRot. There's a bunch of papers that, that tell us that after, uh, like 30,000 tokens for most LLMs, the performance really becomes bad. Uh, in fact, sometimes less than 50% of like the intelligence that the LLM has. Um, and that's bad. So you have to optimize, like despite having the long context to context lens, you have to make sure you're giving it the right context to answer a certain question. On the other hand, um, right now, like, so that's something that already exists but needs to be better. But right now there's very few apps that are personalized to like AI personalization in any way right now. So, you know, when the internet started, there was like after Web 2, everything started becoming personalized, and that was the essence of the new, new web. Um, with, with ChatGPT now being personalized, Claude now being personalized, but everything else not being personalized, it's a big problem. And it's either too expensive to do it or it's too difficult to do it. for most companies to add that to their app, so we provide that as well.

Amazing. Should we go outside?

Yeah, let's do

it. Let's walk Dolores Park.

Yes.

I noticed that you have an ODF poster.

Yes.

Were you a part of— so you were part of BuildSpace and then you're also part of ODF?

Yeah.

So you mentioned that you are a solo founder?

Yeah.

Were you always a solo founder and why did you decide to be a founder?

So I had a few quote unquote co-founders that I was trying out with throughout the journey because Super Memory has now existed for 3 years. And then in all times it either became like, okay, they are doing it part-time and they are amazing engineers and I would want to do it full-time with them. Um, but they want to go to college, for example. Um, or it was like, okay, this guy is going in a completely different direction than what I want to do. So after all of these questions, I was like, I, I already had been building this product for a year, for 2 years at that point. Yeah. So I was like, I'll just keep going and I'll see if I get a co-founder, I'll get it.

Yeah.

Otherwise it's fine, I'll continue building solo. And I stopped looking for one, and I started focusing on the product. And that's when it became like, you know, really like I I think there's a lot of advantages to it.

Yeah.

I discovered them, and I was like, okay, maybe I I shouldn't have went in and looked for one in the first place.

Wow. And so the reason why we're even like chatting with you too is because so Trust Claw was using Open Claw.

Yeah.

And so the OpenCLAW pattern for memory, and like, so, and the users, they started complaining about it wasn't good enough.

Yeah.

So we switched to SuperMemory and saw improvements. So how does SuperMemory deliver better performance compared to OpenCLAW?

Yes.

An existing memory system.

There's a lot of reasons for this, and I'll start with our own core memory engine itself.

Okay.

So the way we built SuperMemory and the way it understands users, We make sure that it's very good across 6 facets. One is kind of knowledge extraction and understanding about what to learn and what not to learn about the user. Yeah. Then there's temporal reasoning, multi-session. Temporal is through time, like understanding the change in time. Multi-session, which is if there's a lot of things that need to be answered with knowledge from multiple different places, it should be good at that. And understanding preferences, user actions, and understanding what even the assistant says. So that's one facet to it, like that's one part to it, which is our core memory engine is really good at these things. Okay. Then there's forgetfulness, which we think is very important, and you cannot get that out of Markdown files. And so you'll have to explicitly ask it to forget or explicitly ask it to remember. And the thing is, you don't know what the agent wants to remember at ingestion time because you can bring up anything in the future. So how do you make it such that it knows everything at all times? So on one hand, you cannot extract everything into memories or save everything. And on the other hand, you have to still be selective and handle updates and stuff like that. Mm-hmm. So the way we do it is we call this hybrid search, where there's both memories that show up as context to the user and the raw information. If the memory does not exist, it will just show up the raw part of the context where it should answer the question from.

Oh, okay, okay. So you have these update edges, so green for new memories and purple for updates.

That's right.

Why does that matter? Yes, that's right.

So Knowledge updates are extremely overlooked when people think about memory. Mm-hmm. Because this, and this is one of the biggest problems with OpenCLAW, that sometimes you would just want it to like, you know, invalidate some previous piece of information and have a new piece that's built on top of it. And so how we do it is we call them like, you know, the update edges. where that old fact is completely forgotten. It's not completely forgotten, it can still be brought up, but it will— the agent will know that, okay, this was the prior one and this is the newer one.

Yeah.

And the new one is built on top of it.

So what's the difference between super memory and just jumping everything into an agent's context window?

Yes, this brings me back to the point I said before, which is Dumping will work, but it will make your agent worse in quality and worse at just answering the question in general. So, and also there's a limit to how much you can dump. You still have to be selective about what exactly you want to give it. And OpenCLaw does this really annoying thing where it will only send back the 15 last messages to, or it's configurable, but like, So imagine you have the last 15 messages with all the tool calls in between, and then you have the memory files below it. So yeah, like you'll spend too much money on the tokens. It will be slower to respond every single time because the way LLMs work, because this is not prompt cacheable because of the compaction. Yeah. You will be sending the entire thread to the model provider and get it back. On every single turn, so you'll be billed like crazy. So yeah,

okay, so it's just more like organized.

Yeah, it's— so in this case, you— it's prompt cacheable because—

What do you mean by prompt cacheable again?

So the way SuperMemphis returns the memories, it can be cached by the LLMs as long— and the, like, our SDK is injected at the right part of the messages array where it will, you know, like, the static information is on the top And it rarely changes, so it's prompt cached.

Yeah.

And the dynamic is in the bottom, and we only update the last point of the dynamic in the profile.

Okay.

So maximum, like, you know, 99% of it is actually already sent to the LLM before, so you're not billed on that.

Yeah. How does Composeo now fit into SuperMemory's architecture?

Yes, uh, this is awesome because we reached out to Composeo last year.

Okay.

Uh, asking if they want to use us.

Okay.

And now we use you guys as well. So we have a consumer-facing app where, you know, it's also an agent and you can use any LLM with it. It's the same thing that I mentioned which I built for the first version of SuperMemory.

Okay.

But now the point is that you can take your memory and you can go to any LLM you want. And so we were like, okay, what tools should we add to make it like really good? Like OpenCLAW has, it can do anything. We were like, okay, we first added the linear tools, the Google Drive, we added all of these tools and then Soham from Composer told me about Toolbox. Oh, hello.

Hello.

That's the amazing clip. So Soham told me about Composer Toolbox. Wow.

And I

was like, okay, I'll just add Toolbox and it has everything. So now the only 2 things that SuperMili can do is use Composer Toolbox, which is do anything.

Yeah.

Or reference its memory, which is awesome.

Nice. Uh, what would building those integrations yourself would've looked like? How much time would it have taken? More time?

Yeah, it would definitely have taken a lot of time.

Wow.

Um, we also have this Sync product, uh, right now, which we did build ourselves, uh, which syncs from Google Drive. Like, the reason for building ourselves is like very deep, but it syncs from Google Drive, Notion, web pages, all of these things. S3 and all of these things. And it was such a pain to build it because like we have to manage auth webhooks to make sure that, okay, everything is in sync. We have to debounce updates by 10 minutes. So we have all of these complicated like workflow logic and we had to do that for our enterprise customers, but we wish that we could use Compose for that also.

Oh, okay. So you mentioned enterprise. So are you targeting— are your consumers— like, is it consumers and enterprise? Yes.

So there's 2 products. Like, Supermemories is many companies in one company, but the 2 big products is the consumer-facing offering, which is our app, which is your memory wallet that you can carry around anywhere. And then there's our API, which these companies use to add memory to their own chatbots.

Okay. How long did it take to integrate into Compose UI?

Yeah, literally, uh, we thought about it like in one of these morning calls.

Yeah.

And we had a PR done in like 45 minutes and merged in an hour.

Oh my gosh. And actually, side note, you mentioned how you were at the Compose UI office and you knew the team. How did that— how did you know the team? What was the story behind that?

Uh, I have known the team for very long now. Okay. I actually

don't remember how I met them.

Um, but there was a time when we did not have an office in SF and we just worked out of Composio office.

Okay.

And, uh, there were only 3 people in the Composio office and then there were 3 of us at the time. So it was like very early days for us.

Oh my gosh, early days for us and early days for Composio.

Yeah, very early days for everyone.

Where were you before that? Where were you living before moving?

I was In Arizona, I was studying. I dropped out to to start this company.

Yeah. Arizona. Wow. And so, did you take a big risk dropping out? Like, did you feel like it was a big risk for you?

Honestly, I would say that I also dropped out of high school, as I mentioned before, to start company.

Oh my god!

And then I sold that company, and then I started another one. So this time it felt like all right. Like no matter what, I'll be fine.

Yeah.

So that's why.

And it's it's fine.

Oh, I now remember how I know Compose Your Folks.

How?

They reached out to me because they wanted to hire me. Oh, at the time I was like gonna go to Anthropic. Yeah, but then I decided to not do either of those. So

yeah, and so because I think another big thing, or what people are nervous about, is they feel like they need more of that startup experience. They need to work at a startup, they need to work at big tech to gain that experience, mentorship. And you just, you dropped out of school, you just went into it.

I would say you still do need it. Okay, I was very lucky to get, get the both the startup experience and the big tech very early because I was at Cloudflare. So while I was building the version of SuperMemory, I was working at Cloudflare. I was working on like infrastructure and research, but also then I was doing developer marketing there. So yeah, like, um—

And you were, you were how old? 18. 18! How did you get that internship?

So this is an interesting story, but I was building a lot of like— right after coming to the US, it was fresh after the acquisition, so I had no purpose in life. So I was like building a lot of things.

Yeah.

And I was tweeting about how building is so easy today, like with all of these things. Yeah. And AI was not a thing yet, so it was just like, with my stack you can build anything real quick. And, uh, the, this, I made like a template and that also went viral. And, um, so essentially in one of these products that I built, it went like so viral, it got like 500K views. in like a few hours and then a few million later. Oh my gosh. That people started attacking the platform in any way they could, and they actually were able to DDoS it. They were able to publish like a lot of like, you know, very interesting things despite me having all of these filters. And so I had to reverse engineer Cloudflare to change the rule sets API and like to do all of these things to protect myself against the DDoS. And I wrote a blog about it, I made a video about it, And the CTO saw it, the now CTO. So at the time he was VP of Engineering. He saw it and he just, he was just like, do you want to work at Cloudflare? I said yes. I got the offer letter in a few hours. So yeah.

So amazing. Oh my gosh. Such a great story.

Oh my

gosh. Yeah. Yeah. Bye. Yeah, bye.

I love being here though. I think like you should always be around kids and dogs. Kids and dogs. Yeah. Yeah, that's that is the way.

Why? Why do you? Because why kids have exude this like youthful energy and they're very like.

Yeah, they're pretty much never sad and if they are sad, they're sad in a cute way.

Oh yeah.

So yeah. Yeah, but I was just— so first I was writing about blogs.

Okay.

On this blogging platform.

What blogging platform?

dev.to. Oh,

okay.

Have you heard of it?

No.

Yeah, it was just like some corner of the internet.

Wow.

And, um, even there I was just sharing like, oh, this is a little CLI I made, oh, this is something I made. And over there no one saw the posts. There was one of them that got to like 500K views. And like, that was funny because it was about Docker and that was my entry into Twitter. So yeah.

How was that an entry into Twitter?

So like the follower— the founder of Docker and the current CEO of Docker both followed me. So I just mentioned like, oh wow, like shit, like both of these followed me. They were one of my first followers. So that's awesome.

Wait, what year was this?

That was 2020.

2020. And you were— so you were on Twitter, that you started going on Twitter?

Yeah.

Because I feel like— don't you think, like, I feel like now Twitter is becoming more of a, like, a space for founders and like the tech people? Yeah. When were you— when you were on Twitter back then, like, was it different?

Like, yeah, yeah, it was. Yeah, I was not in these bubbles at all.

Okay.

In fact, I was not scrolling Twitter all day, I was just posting things. Now it's the opposite, unfortunately.

What do you mean? Oh, now you don't?

No? I still post, but I still scroll more than I post, so.

Okay, yeah, where you gain the inspiration. Yeah, we'll go back to Composeo. And so if you remember, can you give me a specific example of something Composeo made possible that would have been a nightmare to build yourself? I think you probably mentioned this, but yeah.

Yes, so I mentioned this before, which is the tool router of LED. You know, we wanted our agent, which is Nova on the consumer side of things, we wanted it to feel like It can truly do anything, and it knows a lot about you. So we wanted like a few use cases with that. Like, for example, look at the tasks today and send my team a small you know like nudge to do them.

Yeah.

And so first it knows who my team is through super memory. It's like okay, this is the general tasks of the company. This is the things that. going on. This is the teammates. This is the last few calls that Super Memory synced through Google Drive. So it would take them and then it would, like, you know, use Slack API to send them a message. So it can do that right now, which we— it would have been a huge pain for us to build.

Wow. So what's the world you're trying to create with Super Memory? If this succeeds, what does that look like?

So I think in like a post-AGI world, right, like we'll have robots walking around serving people, like anything you can imagine.

Wow.

You have to imagine it highly personalized and you have to imagine that it knows, like there's a feeling of, okay, it's truly extremely smart. Like even with OpenCLAW, like the biggest problem that people had with this new technology is that it's memory. And there's a reason for this. Like people Like, people want memory to be the building block. I think like agents is pretty much solved or is being solved, but memory is the missing component to it which we want to enable.

That's awesome. If another founder is building AI products right now, what advice would you give them?

So I always say you have to have— you have to be really public about it because to do anything new you have to get like a lot of people to believe in you. Okay. You have to approach it like, I think infrastructure is the better place to be, which Compose is also building. But I also think that there's a lot of consumer use cases that are not enabled. So you have to have like a better product sense. And for that, the only way to do it is do the build space way, which is talk to users, et cetera, et cetera. So yeah, like, and I think like people are not doing as much research as they should be doing with all the benchmarks available, like with everything literally open source. Like you can just find out how you can improve it. And there's so many low-hanging fruits in all of these agents that I would want, like all of these are very technical, but I think that's pretty important to build a good product.

Okay, and to really know the technical and like— Yep. Do you use like Perplexity for research?

No, I just use Claude for everything.

Oh, Claude, okay, for everything. And you mentioned talking to a lot of users, like how many users did you have to talk to before you realized like this was like a viable, like a good thing?

So the interesting part is SuperMemory was never supposed to be a company. It was just like this build space project I was building while I was at Cloudflare. And so what happened was I became a consultant first. Because I was posting about like, oh, I have now have 100,000 users on this app, but I only have to pay $5 to run this. And all of these companies who wanted the same, they were like, dude, we pay like fucking $500 for this. And so I was like, oh, let me come help you. And I helped them for whatever use case they had. And, you know, they also like, I made them really happy. And I worked with like more than 30 companies. And one of them actually raised from a16z. And when they were raising, they said that, oh, like in their Series A, they said that, dude, like our entire infrastructure was helped by this 19-year-old. And that's why a16z reached to me. And then that's when I started it as a company. So yeah.

Wow. Wait, is a16z— No, I

ended up not raising from them. It was Speedrun and they wanted me to do the consumer aspect of it. So yeah.

So who did you raise— how— who did you raise from?

I raised from Souza Ventures, Browder Capital, SF1, which is Sudarshan's fund. There's, uh, it's like a really cool guy. Um, and Jeff Dean, who's the head of like research at Google, and like a few other people at Google, Microsoft, OpenAI, Meta.

Wow.

Yeah, just like researchers everywhere. So

How does your team look like now? How many people are working on it? Who are they?

Yes, so the way I like to think about the team is I like believing in underdogs. And so right now the team is mostly people who are low, like less than 25 years of age. That's pretty cool. We have 2 researchers, 1 product, 1 like one engineering and one growth. So I mean, everyone is an engineer, but yeah, that's how we think of it. So 6 of us now all work in the office. Uh, it's really fun people to be with, and yeah, like really smart people. And I met all of them years ago, like all of— I have known most of them for more than 3 years.

So yeah, that's the best. You've known them first, because some people say that you can't— you shouldn't do Business with friends.

Yeah.

But it's been working for you and you wouldn't have it any other way. Like, you think, like, how does conflict— like, when you're— when you're— you have conflicts, how do you resolve those? How does that work?

Yeah, so we, we usually don't have conflicts because we are all pretty aligned. Because the way I met these people was also because they started contributing to SuperMemory, the open source version. Um, and I was like, dude, like, okay, you have been contributing so much for years and like Just come and I'm building a company now. So I had to poach them from wherever they were working already.

Convince them.

Wow. Yeah, that was really nice. And you know, like it's pretty hard to hire, especially people you already know and get them aligned to what you're building. And I think that's pretty important. So you have to essentially become a salesman. A what? A salesman. You have to do sales.

Yeah, you gotta do sales. Yeah.

Like you have to do sales to even get people in your team. And like once they are pretty aligned, then there's not that many conflicts to begin with because everyone is like, okay, this is the basics that we all believe in. And now for anything that like, you know, what's the literally the better solution here? Like what does the data say and what do the people say?

Yeah, wow, that's amazing. And you're going to continue growing it? Are you hoping to relocate to another office in San Francisco or? Yeah, we are.

We are. We are looking at a few places. Okay.

Nice.

Because I'm like I mentioned after the next round we'll have like a few more people so that that will become too too small of a space.

Yeah.

So yeah.

That is amazing. Okay. Any any last encouraging words? Do you have anything you're excited for in the next? You mentioned how you're going to be raising for your Series A.

Yeah yeah yeah.

After, uh, so when did you decide, like, when do you, when do you know you're ready for that?

Yeah, so I think, like, I'm a huge believer in these inflection points throughout the world. Uh, in fact, like, you know, I, I, I was watching the first rocket, like the SpaceX rocket, like land, uh, when I was a kid. I was like, okay, that was an inflection point for spaceflight. There's, there was certain inflection points for LLM inference. And there's been all these inflection points throughout humanity, and like the latest one was agents with Claude Code and OpenCLaw.

Yeah.

And now with the inflection points with agents, you know, like happening, the next one is memory. And so the market is like, you know, asking for stuff like this pretty much every single day. And we are currently able to say that, okay, we are one of the better solutions for memory, and we have like good revenue, we have good Pretty much everything you would need, and yeah, I mean that's part of the reasons that makes me really excited for for what's next.

That's so amazing. Thank you so much, Ruvia, for chatting with me. I'm really excited to see where Super Memory is going to go, and where should people check you out? Where are you most active on?

I am the most active is probably my Twitter, Super Memory Twitter, our Discord server if you are really interested in the product. Are you hiring? I am hiring for a few roles, both in infrastructure. Like, the combination of infrastructure and research is one role, and another is developer relations. So if you are like, you know, really good at building products from 0 to 1, then reach out to me.

Are you selling this merch?

Oh yeah.

How does it—

I'm not selling it, I'm just giving it away. Oh, one important thing in the back.

Unforgettable. He is unforgettable.

Exactly. And this is—

okay, yay, thanks.