Allie Howe All right, welcome to the Insecure Agents podcast. My name is Allie Howe, and I'm here with Cailyn Yong, founder of Momo. Cailyn is in Founders Inc with me, and that's where we're shooting today. Cailyn, would you like to introduce yourself and tell us a little bit more about your work?

Cailyn Yong Yeah, hi, I'm Cailyn and I'm building Momo. Momo is a web app that keeps track of what teams are working on everyday by building a company memory and also integrating with your most used apps like Gmail, Slack, Notion, GitHub etcetera.

Allie Howe Yeah, amazing. That's super cool. I've been seeing you blowing up all over Twitter. So it seems like get Momo, sorry is like gathering so much steam, which is so exciting. I know we both started Artifact, this program at Founders Inc like 3 weeks ago. What sort of follower counter metrics did you come in with and what does that look like now 'cause I feel like you've just completely. Taken it away.

Cailyn Yong So start of artifact. I already had a couple of followers, not a couple, but like I had like 3500 because I sort of went viral with the previous post that I uploaded before. It wasn't related to the product flow. And then I launched last Thursday and that post went pretty viral. It got 1000 likes and then 67 K views and then that sort of also increased my follower count to now 4. 8 K. And then I launched another video this Monday then that did better than the first one. So that got around like 1. 4 K likes and then it had generated 120K refreshes ish.

Allie Howe That's super exciting, huge Congrats. That's super cool. I know like the field of personal agents is really sort of like taking off and notoriety because of like the Claude bot thing that just happened. Do you feel like that's sort of added to the momentum?

Cailyn Yong Yes, I intentionally to ride that momentum. My marketing tactic was like Clawbot for teams and then that sort of took off. But The thing is like, I think personal assistants have been taking off since last year and there's been a lot of teams that have went really viral on X, for example, like Poke, everyone used it, tried it. There's been so many similar tools to personal assistants like texting your agent over WhatsApp, Telegram like there. There's been so many use cases. I can't even like list them because I, I can think of like 5 teams at least that I've been building on that. And I was also one of the teams that were building personal assistants before I pivoted to what I'm building right now.

So it's been like a year journey for me now. And then I knew there were some like set limitations to building personal assistants. I also have like a lot to say about that. But like personal assistants go really viral because it spares a lot of like curiosity, like you're very interested in like, oh, you can actually just like talk and text with the agents and they'll do everything for you. You don't need to go to the calendar or whatever. So at first glance it looks like really cool, but then after a few days, people start to drop out. They don't use it and they stop using it.

I think one of the main problems with personal assistants right now is like retention and trying to grab people to make it a daily habit of like using it everyday. I'm sure there are power users, but but yeah, with if there are not like the right use cases or if it's very difficult to install, then it's I think like that's where a lot of people drop off and start not to use it. That's also why I decided to pivot to a larger problem like company memory.

Allie Howe That makes sense. OK. Yeah, that's super cool. Is it challenging to sell this type of product like to companies versus consumers that maybe have like less security concerns?

Cailyn Yong So right now I'm targeting teams and then the marketing is thought for teams, but the people that have like signed up on the platform and started using it are actually using it for more personal use. So now I'm like, should I change my marketing? Should I also like tailor to personal needs right now? So that's something that I'm thinking because I feel like if you want to onboard teams, you also need to convince the individual user first in order for them to invite their teammates. So I'm trying to get really good at these specific use cases where one person might use it like really well for themselves, and then that convinces them enough to invite other teammates on their team.

Allie Howe That makes sense. I think what happens with personal agents, if I was to assume, is that they're so powerful, there's so many things that they can do, and so maybe users get like overwhelmed and not sure what to do.

Cailyn Yong Yeah, I think that's one of the like the issues with Clawbot or like personal, my assistants is that you can connect to so many tools and there's so many stuff that you can do and there's too much liberty on what you can do. So people don't really know what they need to do next. Also, if you're non-technical, then that's another like barrier that you have to come across. You can't like I can't imagine my mom like installing MCP's on like she's never going to do that. So it's really about the user experience and then getting the right like target audience and getting the right specific use cases for them to use and setting up a template for them so that they can just use it without having to set up everything.

So for Momo, I specifically chose the work like teamwork area because I knew that when you're doing too much liberty, you don't really know what to do. So I'm targeting people to just like connect anything about their work decisions. So it's gonna be really focused on their work, their business, not like personal life or like whatever, just like the things that you did today, the work that you did today. And then also for the teammates as well so that you guys can just focus on the company memory, not like all the other stuff that are like both. Yeah.

Allie Howe And like, memory is such a hard problem to tackle from like both an engineering standpoint and also security. Because like once your memory context is poisoned, if you've got untrusted content that's coming into like the memory, that can create a whole bunch of like issues with like prompt injection and stuff, but also like long running stateful agents having memory over time. Now you're getting into problems like compaction and stateful agents. And what are those engineering challenges like, and how are you solving them?

Cailyn Yong Well, one of the main challenges right now is that I'm building a self evolving memory. So once you connect to your Gmail, GitHub or Slack, they intake all the work decisions that you've made within the platform and it starts adding the memory nodes onto the memory graph that is built. The thing is like after a couple of weeks you might not really need the outdated data, so it also needs to auto purge the old data or it needs to update the old data. For example, if I was working on project A and then now the team is working on project B, then like and if they are too connected, project A needs to be updated so that it can refer to the project B.

But because this is all like I'm trying to build this like auto evolving those kind of stuff are pretty hard when you need to set up all the logic for it and how to update or how to delete the irrelevant sources. And everyone has different thoughts on what's important for them. And when it becomes a shared memory, then that also needs to take care of like the blockers, what's more important, what's that's important, those kind of stuff. So that's like 1 main thing that's difficult to solve in terms of.

Allie Howe Yeah, it makes sense. It makes sense. How is Momo similar, different than Claude? And what can I do with Claude?

Cailyn Yong I think the main difference is that you can build a shared team workspace and shared team memory. For Claude right now, it does have memory, but it's not like installed with it. You have to manually add the memories or you have to install MCP's to keep track of those memories. But for Momo, how it works is that it automatically keeps track of those memories and then it builds it for yourself. So that's like one thing. And then once you have your specific own individual, like team work memory set up, you can share it with your other team members. So that's the biggest difference so that you can just like, yeah, I think sharing itself is the biggest difference.

Allie Howe So that's really cool. I was also wondering, how are you thinking about AI security? Have you seen anything like prompt injections come through or faced any sort of challenges like that?

Cailyn Yong Yeah, this has been an ongoing challenge that I've been dealing with for the past year. So right after I launched Momo last Thursday, there were a couple of attempts for prompt injection, and they were kind enough to tell me that they tried it out and then it failed. So I guess like Momo successfully prevented it from happening. That's because mainly like I knew that this would happen and I put up some like safety guardrails so that they wouldn't be like allowed to do the protection techniques. There's also some well known protection techniques that are on the Internet. So you can just like scrape them and then train it to the LLM or filter their filter them out through like text classification before it even runs through the LLM so that it just like blocks it and it never runs.

So that's like one way I'm doing it. The other thing is like database you like, you need to secure your database a lot if you don't put up the security levels, like anyone with the public anon key or the public URL key can, yeah, can see or like retrieve other people's data within the database. So you need to put up the right like security guardrails for that as well.

Allie Howe Yeah, that makes sense. Sounds like you're doing all the right things. That's amazing also that that happened like so soon. Yeah, I think a lot of people I talk to things like, oh, like this will like never happen to me or like people aren't, you know, smart enough to be like, you know, doing AI exploits and stuff. And it's like, no, like it it it's happening. Yeah. If you want to run like an agent, like in production or like build in public, like you should expect it. So super cool that you were able to tackle that. One thing I was seeing with like the Clawbot that was happening was like the skills hub they built for it.

Someone made this skill. What would Elon do? I think it was called and they like hacked the skills hub such that their skill show the most number of downloads. They just like put like a full or something on like it the most and.

Cailyn Yong I see. OK.

Allie Howe And that's like one of the ways like we as users trust, like, hey, like can I download this MCP's? Yeah. Can I use the skill? Can I even use this like npm package? Like which one has the most downloads? But if you can like pack that or get a way around that, that's like the easiest way to like exploit like user.

Cailyn Yong Yeah, even like MCP's, they're not fully secured and you can't really know if they're secured or not. So you also need to be really careful about installing all these MCP's on the Clawbot and giving just access to everything about your local files and everything. So there's a big like security risk to Downing all of that.

Allie Howe Is it a? Feature in Momo to be able to like, share skills or connect things to Momo or how are you thinking about that for your product?

Cailyn Yong No, not yet. One of the people in the comments have asked me if they could just like hook up some API about books and then extract it. The thing is like right now, first of all, like I don't trust MCP's and I don't really want to use MCP's to build memories because right now in Momo we have our own like memory extraction logic in there. So it doesn't really make sense to use MCP's to forward that. And then what I was trying to say is that for each like app that you're connecting, there's a different logic to how it's being extracted. So I'm thinking like, like if for me to scale, it would probably be really fast to just like hook up like Zapier or like those like webhooks are in some piece.

But I just wanted to make sure that the quality of the memory extraction is right for each individual app. So right now I decided not to go down that path and just like work on the main apps that people want and then just get the memory extraction logic really right for each of those and then see if we can scale from there. Yeah.

Allie Howe Yeah, that's like a plan. I'm always nervous to like install MCP servers myself for. The same reason. And there's like various like MCP scanners out there on the market today to help you find things like tool misuse and stuff and then run poles and all that. But yeah, definitely it's like a huge, like a wallet build. And it's like if you don't need the MCP server, and especially if you don't need all the tools that come with it.

Cailyn Yong Yeah, yeah, yeah. I can't. Like it's, it's super easy to just build the tools instead of having to like install an MCP again and then use those tools because I hate installing MCP's and there's such a yes, it's like really hard to onboard.

Allie Howe We thought about using like coding agents to like create like the API's connections, like to like the Zapier API for example, and like you do code execution instead of an MCP. Server,

Cailyn Yong yes, I've thought about it, but then you also have to pay for Zapier, so I was like, I'll just build it myself.

Allie Howe There you go. That's awesome. Yeah, I mean the cost of like code creation these. Days is so. Small, We're creating more, more product than you ever have because of AI. But we also have like a code review problem because it's so easy to create all this code. And even as I think both of our team short companies are very small, even given the small size the amount of code that our teams are generating. Is huge. And so I think one of the best ways to help with that code generation problem is like creating like architecture diagrams and being able to say, okay, like given the new code, where does the largest like security concerns lie?

I was going to change my architecture. I'm going to the way the agents deployed. Is there different pipelines or anything you've built in your workflow to be able to help with that code generation problem?

Cailyn Yong Yeah, For us, we really need to keep track of all the pipelines that are happening behind the memory layer logic. We want to know like how the decisions are first being extracted and then how it's being updated to the memory graph and then if it's correct or not, and if it's being encrypted, for example. So there's a lot of steps that go through before the memory graph itself is created when you're first connected now. So we needed to keep track of all of that process. So for us internally, we built a pipelines Argo to show how like how these steps work. And then, yeah. And then we were keeping track of that, the token cost, whatever, like anything that you need to keep track of.

Allie Howe Yeah, that's awesome. And I know, right, because we had you try my tool Verde, which creates architecture diagrams and it labels where the lethal trifecta exists within your architecture diagram. So you can start to understand. Like how do I keep the different pillars of the lethal trifecta, private data, external communication, and untrusted content away from each other such that you prevent AI exploits? What was that process like for you? And is that diagram was that helpful? How is that different or similar to the diagrams you create with pipelines?

Cailyn Yong I think it is helpful because just today what I was going to do was go through with each of the apps that I'm connecting to Momo and see if they're being properly encrypted, if there's any security leaks, if there might be any vulnerabilities for prompt injection or hacking, those kind of stuff. So I think it might be helpful for me if, if I could, yeah, extract like individual pipeline diagrams of each of those apps, like each of the functions. Because I think for the entire diagram of what's happening, I probably know what's happening. But for like the specific functions of what's like happening inside, I would like to see a more like specific view on that.

But the diagram itself is it gives like a really clearview of what's happening and what kind of security risks that I might have, which I'm dealing with all by the way. So, so yeah, it's like it's, it really helps you all.

Allie Howe Super cool. Yeah, that's great to hear. And yeah, so that'd be super cool to like zoom in to like different. Parts of that,

Cailyn Yong yeah, yeah. And let's see like what's happening within like each of those like.

Allie Howe Yeah, that'd be super cool. What is next for Momo?

Cailyn Yong Good question So I watched the social Network 2 days ago and then it caught me really thinking what should I do to make the next billion dollar company? What are the right decisions? And I've been just like thinking about this every single day. What are the right decisions that I need to make right now in order to progress, look further and make bigger streams and bigger outputs? So right now, I haven't really finalized this yet, but if you think about Slack or Notion, it's really about Slack. One liner is how teams communicate. Notion is for how teams keep records of their documents. So for Momo, I think I also want to do where Momo is like a workspace where it shows how decisions are made and evolved over time.

So it's sort of like a truth clear that you check every time when you start work. So before even like hopping onto Slack or Notion or any of the workspace tools that you use, you just like check this first and see what's just happening overall in your team and your company, The sales progress, the products progress, and then you get a overview of what's happening within your day, within your week, and you move on to the next applications. I think like that's something that he wants to build just like something that everyone uses every single day. So I'm like thinking in those lines. I'm not sure if I have a great like one liner for this yet though.

Allie Howe I think that plan is is amazing. It'll be super cool to see this get plugged into more people's like workflows. And I think it'll just be this viral thing of like, oh, you know, I'm user X and like I posted about how this, you know, 1 workflow. This one use case saved me so much time every day. I can't imagine my life now without this. Personal Assistant. Once that like catches on, it's just going to go like wildfire, right?

Cailyn Yong Yeah. Yeah, there's like, yeah, there's a lot of issues with like, Slack as well. Like there's like too many communications that are happening and then people just can't keep track of, like, what's happening and what they need to see or what they forgot to reply. That's also something that I'm interested in because Momo keeps track of like all the decisions that were made within Slack, and then you can keep track of that within like, time periods and also by which team member made that decision. So I think it might be pretty interesting to see how this plays out. Maybe it could be a new version of Slack or AI native Slack, but I think, yeah, there could be so many different ways that this could go right now.

It could also be a new AI native Salesforce if it keeps track of the clients memories. So there's yeah, I'm trying to put my finger on. What's the next step to take?

Allie Howe Well, I've got absolutely no doubt that you will figure it out. You are absolutely on fire. Like you're a huge inspiration to me, honestly. Super cool to see that this is taking off so much. And thank you so much for coming on and taking the time to tell us more about Momo and the security challenges and what it's like to actually build an agent, take it from a proof of concept and actually ship it and actually make it work. It's a huge difference there. I hope we chat again soon.

Cailyn Yong Thank you. So much thank you so much for having me.