Solving Context Loss: The $20M Problem Every Startup Has

Why not let AI do the job? Being a second brain for the teams tinker is like your second brain to your team. and the thing that we want solve is basically the context loss for teams.

Is there any similar, tools on the market right now?

Honestly speaking, I think there are a lot of great tools, for AI being an agent and doing all these work for people, but we have not seen any chat tools

that has as complicated AI memory structure as we have built. But in the future, what we are aiming to do, what we are envisioning to do is that AI can give context to human. How do you see tener in three years? We definitely will redefine how.

Hey everyone. welcome back to ai. Today I'm sitting down with Kisan,of Tener ai, who's taking on one of the biggest data problem, in AI to solve, which is memory and context. Kisan, thanks for joining us. before we dive into Teka, could you tell us a little bit about yourself and where you're [00:01:00] working on?

Yeah, sure. hi everyone. thanks Angelina for, inviting me to this podcast. Very glad to be here. I'm Kisen. I'm the CEO of Tonka ai. And prior to Tonka ai, I was also the co-founder of Mind versus ai, where we built, an AI agent platform and long-term memory for individuals. And at Tena we built long-term memory for teams, using a tool, which is like Slack plus Glean plus mena.

this AI agent can execute on your team's behalf and it grows with your team. So, we say that this is your AI co-founder. and prior to starting my, AI entrepreneur, journey, I was a strategy director at TikTok. And even before that, a structure strategy manager at, meta. so I've been roughly, meta and then TikTok for around eight years and been an entrepreneur for four years.

that's an amazing journey. I mean from managing strategies and product to building a product yourself and building something related to memory for individuals and then to teams. So that seems a natural transition. Can you tell us a little bit [00:02:00] more about 10 Care, it sounds to me like a second brain type of product.

are you targeting.

at Founders. Yep. so Tinker is like the second brain to your team. and the thing that we wanna solve is basically the context loss for teams.

for a lot of teams, no matter big or small, one of the biggest. Pain points for them is context loss.

people are juggling amongst a lot of, SaaS tools or applications. and when you're juggling amongst them, you have to understand your, chat history in Slack and then summarize it for instance, into , notion,

And another context loss scenario actually comes from employees come and go. So when you have one employee leaving your company, they actually take the context with them.

and handover document is always nightmare. So is the onboarding document. When you have a new employee on board, you don't have enough time to update the onboarding document all the time. So usually, especially for fast moving companies, they are just like, you can learn on the job, compare about yourself, right?

Yeah. Yeah. And that brings a lot of pain to [00:03:00] both the employee and the company, the managers. so employee onboarding and offboarding another context loss actually comes from implicit context. We have everything in our mind, but we don't necessarily have the time to communicate it. Out loud to everyone within the company.

And so gradually throughout the, the years, people start having different contexts. They have like partial contexts and it's very hard for people to get truly aligned. and we thought that why not let AI do the job being a second brain for the teams and, If AI can remember everything and brings to you the right context, the right information when needed, it solves a lot of problems that we mentioned just now. and the key is basically let AI see all the context. So we built this team messenger,like Slack, but it's connected to your notion, Google Docs email, WhatsApp, et cetera.

So it gets the context from your chat history and also from your, emails and documents and, it automatically updates itself.[00:04:00]

So why chat? the reason is that actually, a lot of decisions are made in chat,

So when we have information and reason and decisions, AI will be able to understand how this not only what's going on in this team every day, but also how this team makes decisions. that's the full context that we want AI to know.

You mentioned about three type of, context losses, one of the three was, implicit context. And did you mean that things that we are thinking in our mind but we haven't communicated?

Yeah. Something in our mind haven't communicated. Um, so, so that, how let yeah.

Know about that, what I'm thinking.

actually you'll be so surprised at, how, how consistent our thinking patterns are. Really So meaning that based

on the context?

Yeah, based on the context and also based on our historical decision making. so one of the features that we have is deep thinking, reply, smart reply to your messages.

when, you know you have an incoming [00:05:00] message and or a group chat and you can just swipe up and the AI reply pops up. And the AI reply is with, Deep thinking. Basically it's with the context within the chat window, but it's also with the examples of your previous replies.

and it's also taking into consideration your position in the company. Whether you're a CEO or a product manager, or an engineer, like your role and responsibility, and also taking into consideration your company's, mission, vision and all these, so those come from memory. And then there's few shots learning from your previous reply samples.

And we take all these, we blend all these, and the replies are sometimes like really amazing. so especially when it comes to like decision making.

So sometimes I find myself relying on the, AI smart reply more and more because every time I just swipe up and it gives people the context from other chat group, to this chat group.

And also it gives people the right decision based on my previous [00:06:00] decisions. So that saves me a lot of time. First of all. I don't have to copy and paste the context from one group to another for other group members to understand. and two, the decision makings are quite, consistent with my own decision making, so I don't have to spend time writing down the reasons why I make this decision.

that's what we do. as humans. ' Yep. If some questions are asked in one channel and then it reminds us. Okay. Some similar answers might have happened in another channel. We probably do the copy pasting manually. That's correct. That's correct.

for instance, the other day I was,Discussing about hiring decision with our recruitment manager. And we decided to give an offer to someone, based on these reasons. Andwhen I try to communicate this decision with our team, who this new hire is joining, I just tell my ai, Hey, uh, tell them why we are hiring this person.

And this AI writes all the context. And I just like forward this whole reply to my team.

That's [00:07:00] really nice.

Yeah.

Do you envision, this is for everybody

on

the team?

It's not

the CEO?

Yeah. It's for everybody in the team. Everyone in our team is, is actually using this AI right now.

you mentioned Y chat. I can see that this is the major way of how we communicate with another human being. So now we're just conversing through AI and with our teammates together, right? Yeah. So merging the machine communication.

With the human communication together,

we try to take a first principle thinking, right? A lot of times in team setting, the ideas actually start from you chatting with your teammates. chatting with ideas and notion or other documents actually comes as a second step or third step.

So if you can, if you can capture ideas in the first step, when people are communicating about ideas, when people are communicating about decisions, then you probably understand, what's going on. Then you can make document AI generated documents based on that. But when you do like [00:08:00] documentation, it's probably already too late.

Because when people are trying to write down in their documents, they already have ideas structured in their mind, and they probably have already communicated with someone. Of course, they share documents and communicate, but that's more formal sharing, right? So we wanna capture ideas, we wanna capture decisions, at the time they're made and yeah, top of the, yeah, that makes AI more accurate.

But of course we have API to connect with, notion Google Docs. So we basically have everything including your meeting notes. and this AI will be able to understand what's going on and. How does that impact your decisions?

Are there any similar, tools on the market right now?

Honestly speaking, I think there are a lot of great tools, for AI being an agent and doing all these work for people. when, this year especially, we see a lot of companies talking about, building this AI agents, AI employees for teams like ai, go to market agent ai, product manager agents, et cetera.

But we have not seen any chat tools or [00:09:00] any, AI messengers or a, any SaaS tools that has as complicated AI memory structure as we have built. And the impact of that is,

You have to fine tune the context ., quite a lot. you have to tune the ranking You have to edit it, and sometimes retrieval, is not accurate. So accuracy actually it's very important when it comes to work.

and accuracy takes the right context and the right context comes to. Memory, a well structured AI memory. And so when it comes to team AI memory, we have built the more, the most advanced,infrastructure and framework on that.

since , you're touching upon the architecture of your system, maybe we should go there and then maybe you can tell us a little bit more of how you designed it.

It looks how our team designed it. Our team. Yeah. Yeah. so I got this from your white paper. I read your white paper. It has a lot of information and this amazing diagram and the audience of my channel love our diagrams. So maybe you can also share with us like [00:10:00] what's going on in here.

Right. this is not only the long-term memory. You can see that the long-term memory is only very small part of that. This is a more, thorough. AI decision making engine for our product. Now, before getting into this diagram, I wanna share a bit of context on what we are building.

We are not only building this slack, chat engine, but we are also building the long-term memory where AI can understand what you're doing and what you're deciding on every day. the vision of this tool is eventually AI understands what context to tell you. nowadays SaaS tools is that people tell AI the right context.

People brief AI about the context. We talk about context engineering, right? But our tool today. Allows AI to update the context themselves so users don't have to do context engineering heavily. But in the future, what we are aiming to do, what we are envisioning to do, is that AI can give context to [00:11:00] human, this sound very counterintuitive.

Like, why? Why do I need that? But a lot of times, when human beings make decisions, we don't have the right information on hand. especially when, let's say, when we have a lot of people talking in a team, talking in a group chat. I wanna tell people, this is my decision making.

am I confident enough in my decision? Do I have the right information? Do I memorize all the right information that have, that people have been discussing about, that are available online?

our thesis is that, if people have the right. Information, have the accurate information at the right time, their decision making would be much better, much more accurate, and much more powerful.

So eventually, in the future, we want AI to brief people with the right context. When you're making decision, AI tells you, Hey, in your team, your teammate, a tell once said this thing, your teammate B one, set this thing. So when you're making decision, take those information into account.

And so this decision engine is our, is something that we have [00:12:00] not yet fully built into the tanker. It's still in the lab now. The whole thing combines long-term memory and decision making inference engine andthis whole agent framework. So agents can make execution and everything is based on long-term memory to to be able to give them the context, the right information, and this long-term memory, which is not well drawn in this graph.

It's actually something that we have been engineering for, designing for the past two years when it comes to, team long-term memory, there are a lot of challenges, and I can talk about those challenges later.

But basically we combine a lot of different methodologies, like episodic memory,vector database, knowledge graph, et cetera, for different scenarios.

And a lot of times it comes to how do you optimize the cost versus accuracy because you are using different scenario, different methodologies, at the same time. How do you optimize speed when you're retrieving in different ways? and how do you make the retrieval more accurate when it comes to team environment?

But basically when, things comes in [00:13:00] the information comes in. We use LLM to process, but of course how we prompt and guide the A LLM to extract the right information is something that we have taken a lot of time, optimizing for and after LLM understands the context, then we retrieve the right context, the right information into the right structure.

So for instance, semantic structure or keyword structure, et cetera. And all these are stored in this long-term memory infrastructure.

And now we have, recently upgraded this whole long-term memory infrastructure into an Everman os. so basically the memory becomes persistent.

That's why it's called Everman. and. This memory system will power, all these decision making engine and agents framework, the right context. Now we have built this agent framework as well right now. So you see that in our application, you'll be able to make pitch decks, business plans, PowerPoints, websites, and you don't have to breathe a lot of context.

Sometimes I just say, Hey, make a pitch deck about tanker for, series A [00:14:00] fundraising. AI understands what tanker is doing, right? So that comes from the long term memory, but of course there is the short term memory as well.

and there is the,the, the context, that we put into the execution. so for instance, it's not only, it's not only that we extract everything from this long-term memory, from Symantec understanding, but also sometimes the users want to specify, the information from this memo that I made,from this memo to a pitch deck, make it for me.

and so you have to make that infrastructure ready as well. But those are like engineering work. Right now, that agent framework goes to execution, but our agent goes. Beyond just executing, just making page stacks. Our agent is also in the loop when you're chatting. and our vision towards the future of work is that when people are chatting, agents gets work done within the chat.

so you actually don't have to break down your workflow into several different applications. Nowadays, we chat with our colleagues and we identify. [00:15:00] When we identify a follow up list, we then go to Notion or we go to Jira to write down this, this to-do list. And then we do project management.

So our day is actually fragmented, broken down, into several apps. Totally. But yeah. But we built this agent in the loop so that when you're chatting with people and we talk about a follow up, this agent will just pop up this follow up immediately and this follow up goes to that person's follow up list.

And when that person marks that I've done and this follow up, then. Another person who's subscribed to this task, will be notified. and when the task is done, it goes to that person's weekly work report automatically. So agent takes in all these work, and is in the loop all the time.

So you just have to chat. You just have to, collaborate with your teammates without worrying about juggling around the tools. So that's part of the agent framework as well. when it comes to agent framework, I think, the difficulty is not actually, in the. functional, but also in [00:16:00] identifying the right, movement be to be made.

And that takes a lot of, engineering effort and a lot of data for us to train these agents.

And, with that causal inference engine, AI will be able to understand the, chain of thoughts or tree of thoughts for the users, based on the past decision making logic. And takes the new information from the launch of memory, you'll be able to suggest to people with the new decisions.

Sounds like it's personalized to

each user. So it's, it is trying to understand my decision, tree essentially based on all the context.

Yeah. And also based on your roles and responsibilities.

I feel like you are replacing pretty much all my scattered information siloed during the day-to-day work.

'cause I don't even know if I need to-do list anymore right? Because you are gonna figure out what I need to do. Or even my, uh, team members. Um, yeah, yeah, yeah. That's,

that, that's exactly the goal that we're trying to achieve. like this just happened three days ago.

[00:17:00] one of our colleagues was writing down to herself on a chat, my to-do list today include,I have to finish this PowerPoint, and then she stops while she was trying to think about other to-do lists. AI automatically pops up this, the suggestions to her decision, to her to-do lists.

And she was like, wow, this is so accurate. AI understands all my to-do list, and it comes faster than. I can think about it. And this is already helping a lot of teammates, within our team to, maximize their productivity.

That's super powerful. it sounds to me that the system works across the memory system, especially works across chats and across channels.

chat BT or Clause is not doing right now. 'cause oftentimes, I reach the max maximum of my chats. I'm paying $200, but I'm still reaching it all the time. It's a pain to give all that conversation context to another chat window again. Yeah. And it doesn't, it just, yeah.

It just extra work. So you're saying you're [00:18:00] eliminating that, that boundary of the memory loss. In these transitions.

exactly. models are piling up their contact windows, right? there are eventually gonna be models that with tens of millions of contact, tokens available for the contact window.

But the problem is with transformer structure, it's impossible to have unlimited context versus very accurate retrieval of context. Understanding of context. There's the attention problem. and

so we don't believe that, maximizing the context window is gonna be the ultimate solution. We believe that giving AI the right context is the step one.

Beyond that, you can then add the context window, but the right context. we talk about like better data to better outcome. The right context is the better input.

How do you do that? long term memory retrieval. So you have to understand what's important in what scenario.

and for instance, when people are retrieving, when people ask our ai, hey, the [00:19:00] last time I shared some PDFs on specific information about, Japan market. Can you tell me more about that information? does it find the right contact? Does it find the right PDF? Does it go into the PDF to retrieve the right context?

does it process the whole PDF. when people are sending the PDF. And so whenever PDF that you have that's related to, Japan market will just be thrown into that context window, or do you first of all, index all the right information with the chat window ID and with the document ID so that you understand when people are asking about that specific PDF, you go to that specific PDF first.

so there are a lot of structures to the launch, a memory that you have to design,

right? So how do you make sure you're retrieving the right, contacts for the right task? 'cause you have so many options, especially when the memories accumulates over time, right? You will have a lot of datasounds like you're not, stuffing all the memories into the context window.

So how do you make sure you are accurate?

[00:20:00] that's exactly this whole challenge of, of long-term memory. And I'd say that we've tried a lot of different tools and unfortunately, like all other tools, cannot retrieve the right information, cannot provide, very high accuracy for all scenarios.

Like each one of them can only provide the right kind of like okay-ish accuracy for one or two scenarios, but not all. And sometimes, This information about retrieving, specific details from one PDF as one example. And for this example specifically, we index the documents when it comes into the structure.

And we index the document, as summary and document ID so that we understand.it's basically an index of the library. So basically you understand what. Document talks about what? So you can first of all find the right document, but at the same time, we also index everything into a memory,into a database.

So when people are retrieving about a specific information through, searching through keywords, then we go from [00:21:00] the more traditional way. And there is also the episodic memory when people are asking, Hey, last month I talked about this person a about something about this topic.

Can you retrieve the specific information for me? For instance? can you retrieve,can you find how much is our hotel budget in region A? I asked that question to our HR the last time. So that's more about, combination of episodic memory to find that specific, chat window as well as trying to find the semantic matching.

what do you mean by episodic memory? Um, this is more about, it's kind of like standard memory. first of all, We put timestamp on every piece of memory. because when it comes to teamwork, it's very important for you to understand like when this conversation happens or when this document was written, when this specific matter happened, people need that information to understand, what's more recent.

And you also have to develop this whole thing in sequences so that you understand when people are retrieving this specific information about a [00:22:00] project. It's. Not only about the the more recent progress about this project, but sometimes also need the whole historical progress of this project.

So you can, when you have the time sequence, you can then develop this whole thing. it's also more about this new framework that you are showing here. When you can, if you can scroll down a little bit, this mori that we built earlier on,it's also some episodic memory.

So what we found out is that when you retrieve the memory from a specific conversation or specific story, if you only retrieve,for instance the keywords or a specific, knowledge graph.

sometimes it's not accurate when people are inquiring or retrieving specific information because people, human beings remember stories in episodic ways.

for instance, I've been chatting with Angelina about this. we're doing this podcast right now, and two months later, I might be asking, hey, so the last time I did this podcast with Angelina, what happened,is there any actions that.

that were identified in this podcast.

And so this whole thing is basically a story in our [00:23:00] mind. And this story has its subject and its time and what matter and what happened in this story and what's the implication in the story.

So that's how human beings construct stories. And if we can instruct AI models to remember things in the same way that human beings construct stories, then there is higher likelihood that the question falls into what we have prepared that we ask the AI to prepare.

that's very interesting. So you're adding the time dimension, basically some sort of time-based or longitudinal data as part of the memory as well. What I mean is that if you think about the traditional rag systems, when we are doing retrieval for large amounts of documents, they're very static.

It doesn't constitute a story or a timeline by itself

or association.

Yeah, you're associating everything, including the actions, right? Because the team is operating on their own, speed and cadence. They're doing things, they're consuming large amounts of information, and then they're taking actions.

That [00:24:00] story is not recorded in any way. Nobody's, unless you are recording these in these kind of, sequence. Otherwise, that context it's a lost, we know it as human because we experienced it. That's what you mean by this experiential coherence. Yes.

Am I right?

Exactly. Yeah. And I think, I think when it comes to Monte memory, there are gonna be more and more work that's done, by studying our human brain, like how we remember for instance. the more, recently, used and popular, memory frameworks including hip rag, including metals, ref, rag, et cetera.

If you look into how they make memory, it's quite intuitive, like introducing the structure to our memories and compressing the memories. So it's pretty much like when we are remembering something. There's structure, there's abstract understanding, there's association.

and I think more and more work will be published, will be done in terms of the memory side. And the reason why we are more advanced in memory is because, we have a whole [00:25:00] institution, a research institution on neuroscience.

So they've accumulated decade of experience doing neuroscience research, especially how human brains remember memory. And when we do the memory design, we naturally take this more, human brain inspired, neuroscience inspired, methodology when designing the memory structure. That's amazing.

What do you think is missing in today's, architectures in terms of memory?

I know it's actively researched area for sure. based on your understanding, comparing with human, how human memories work. What do you think is missing in AI today?

I think there are a lot of,similarities in terms of the framework, but when you are putting all these framework in this real usage, you have to, you do have to engineer, engineer the whole thing to make the results better.

And also, how do we combine all these efforts together? there's no one size fits all. I'm not sure if there will ever exist a one size fits all memory in infrastructure to be able to retrieve all,and, comes up with all the most [00:26:00] accurate memory at, with whatever question that you ask, right?

so there's Inherit disadvantage for for lms, for AI versus human brains. Human brains are just like way much more efficient. That's one thing. And also with regards to the transformer structure itself, it has its heart limit, especially the, the attention, limit.

And so I think, maybe there will be other. Innovations with regards to the transformer structure, but with the current structure,there are definitely limitations. and you can only do so much with regards to how you store the information, how you retrieve the information. But you cannot actually, have one size fits all.

So you have to make a lot of frameworks, work together.

And another challenge that we do have is our specific scenario. There are a lot of memory structures when it comes to user AI chats, but that's easy. You only have one user ID and one user chat. With ai, they already put context within the chat window, that's called prompting.

Yeah. when you're [00:27:00] retrieving the memory, it. Actually don't have to do a lot of association. So context continuity is not gonna be a problem for those memory structures.

But when it comes to team chats, you do have to take into consideration first of all, the context, continuity and also the right granularity of retrieving the information.

for instance, people talk about implicit context and people jump topics. They're off topic talks. So how do you associate those topic together? And when people are doing. In a group chat about specific topic, and then all of a sudden someone says, Hey, I suddenly have an idea about this question that we discussed about last week, do you know which discussion that he was, referring to?

Can you make the right association when storing the data? so those are the engineering efforts that we actually have to make. So when it comes to like team memory, it's not only that, it's also about.

Matching the right user id. We connect with Slack notion, Google Doc's email, but users have different names on those,tools.

How can we match the user [00:28:00] ID and how can you match the user ID with their roles and responsibilities and extract the right, the context with the right granularity? Let's say there are some, introvert people, extrovert people. There are managers, they're, there are engineers who only write code and don't talk much in the group chat.

In the document. So when you're retrieving people's follow ups. You actually can over retrieve for,for managers because they communicate a lot to their teammates. And you under retrieve for people who are only writing codes all the time and don't talk much in the group chat.

So how do you optimize the situation? Who do you optimize for? So those are a lot of, there are a lot of, real challenges. And the biggest challenge for all of that is basically your data set. Do you have the right data set to test and optimize the scenarios All the time.

Yeah. Yeah. I can see where the challenges, especially you're saying this is not my individual, memory anymore. This is the whole team. So you have. in a sense, this is like the team God, which knows everything, [00:29:00] knows everybody what they're talking about. And know, knows what's happening within the company and within the team.

Mm-hmm. Mm-hmm. how do you make sure that you give everybody the right amount of context? Exactly like what you said, as a, as a team member, I may not even have the right to know what the managers knows. For the purpose of helping me, whatever the role is to accomplish my tasks, I need a set of memory that works for me.

so how do you balance that? 'cause I think it's very challenging based on what you're talking about. It's a very complex system you're building.

yeah, exactly. It is very challenging. so you have to select who you prioritize for, first of all, we prioritize for founders, for small teams.

So they usually are the decision makers. So you optimize the cases for them. again, there's no one size fits all you, you do have to choose. And two, for scenarios that your AI cap capabilities cannot optimize for, it cannot do it really perfectly.

You have to make it do with it, by having better product designs. so for [00:30:00] instance, when you over retrieve for the founder's case, right? and you under retrieving for, more of the junior members' cases, and you do have to provide a better way and easier way for them to input their follow-ups, to identify their follow-ups.

and also sometimes you do have a lot of follow-ups, on the lists. You have to basically rank those follow-ups. So when people are seeing a lot of items on their lists, they're seeing the most important items. And no matter how many that's, in the expansion list, they're still seeing the most relevant ones.

and that actually don't disturb the user, experience. and there's a lot of, nuances in AI product designs. And another example is how obvious we want a feature to be. Sometimes you have, elegant design like Apple, right?

everything is like swiping and pinching it's very elegant. we'd love everything to be designed in that way, but later on we found out actually the users won't even know what they don't know. and we end up having a lot of [00:31:00] features that the users are just not aware.

For instance, in the chat group, when you swipe right, the group assistant will come out and you can ask the group assistant a lot of questions. And there is like the quick prompting you can do, you can schedule with the assistance. So you can say every day at 8:00 AM summarize to me all my unread messages and emails, but those features are.

Largely unaware by our users because they're so hidden. And and even for the smart reply, you have to do the very simple thing, which is to swipe up. But if a user doesn't know there's a smart reply feature, they wouldn't swipe up. So we are, a lot of times we're, debating between whether we should make one feature, more obvious that the users will.

It's right there. It's very hard for you to not know that it exists, versus making it more implicit. it definitely takes time, right?

Yeah. The product side.

Yeah. And we're also making some of the most frequently used features, more obvious. So [00:32:00] for instance, like replies, instead of you having to swipe up now, it just appears in your window so you can just click send and it's absolutely there.

you cannot mix it. so there are high frequency F usage features that we are trying to make it more obvious.

Makes sense. Yeah. So the market needs education for sure. this is still early stage for communicating with the machine and gets things down. And I do think this is the future

Yeah, but it's very exciting. I see the human to AI interaction or collaboration, not only in the way that AI is with the right context, but also

AI can get things done within the chat window automatically. so there are a lot of opportunities potential when it comes to chat itself.

a lot of things that we can explore.

What about hallucination? Right now? 'cause, we all face this issue, right? yeah. Yeah.

I would say the most, effective way is for you to, limit the AI to find all the information within the team knowledge.

And for every piece of information it retrieves, there is a source that [00:33:00] users can click on and check the source. So either it's a chat ID, or it's a memo. Everything has to come with a source.

Does Tenco work with,mostly unstructured data? Text data or multimodality data. What about structured data?

Do you work with databases?

we do structured data as well. So I would say we don't do like very complex databases yet because we target the small startups and most of the time when it comes to small startups, they use very standardized tools like Notion, Google Docs, and

the database notion, right? So it's not gonna be like too complicated database,Oracle or whatever.

Erp. If you solve, if you do solve that, then this is like a one tool for everything. Seriously. Yeah. Anything else?

Yeah, we try, yeah. We'll try to take it step by step.

so first of all, let's make the standardized tool really accurate first.

Are you handling like role level access right now for privacy reasons?

yes. that's basically designed within the app from day one. Privacy and restrictions [00:34:00] are very important to teams.

so basically,our AI. Only access the data that user, that specific user can access to. And including

the context or the episodic memories as well. Everything, yeah. I'm not gonna know like the CEO, what they did, who they met and things like that.

no. Even the CEO wouldn't be able to access the data that they don't know.

So for instance, if,if your team puts together a group To complain about this company and this group, does not include you. your AI wouldn't be able to read that data.

awesome. Yeah.

Yeah. We're thinking about making a paid feature,for CEOs.

I'm kidding.

The CEO may pay for it. I'm like, I wanna know what my company and my team members are saying bad things about me.

No, we're not gonna

make it. Do you have metrics to track the accuracy levels?

the accuracy is, something that we can track with, some existing benchmarks.

I would say that those benchmarks actually do not care for [00:35:00] the specific scenario that we are targeting. the team chats. And also team collaboration. We do have. A lab within, our parent group. And,this AI lab is gonna make benchmarks,for different scenarios in the future.

we also do look forward to having more specific benchmark for our scenarios so that the tracking will be more accurate. But for now, we just track with our internal use cases. and we basically use our own employees to score any evaluate for the internal use cases.

Examples.

Yeah.

So user feedback, basically based accuracy. Yeah. and also it's not only about tracking the accuracy, the, doing the benchmarking, but also about what kind of dataset you can use. For now, we only use our internal examples, internal chat history as

Examples for data, for evaluation, but sooner or later it's gonna be exhausted, right? Because, our team cannot just produce that many, that many examples. we previously tried to use glaucoma, but it's not really, 100% for the [00:36:00] scenario either. So we do hope that there is gonna be more, specific data set as well as evaluation benchmarks for our scenario.

Awesome. I think how you position Teka is, let's say your AI co-founder,

Or let's say second brain. Do people really understand the category or are you trying to. stand out from your competitors versus how do you balance that?

I would say it's very hard to, for us to join a specific category. I do find it difficult for. to use one word to describe what we do. usually I just say we're Slack plus gle, plus mena. But for a lot of people, they don't even know what GLE and MENA are. if I try to say we're team chats and you can search and your AI can build pitch decks and everything for you.

They're like, oh, wait a minute. So where do I start first? so it's very difficult we position ourselves as your AI co-founder because our audiences are usually founders at early stage, seed or pre a stage, and with five to [00:37:00] 10 people in their teams And that's when they are looking for better AI native tools because when you're already late in stage,

you are probably already on a set of tools that are, more traditional sales. but for newer companies nowadays in this generation, they are more open to exploring AI native tools. And,they're more AI native themselves.

and that I find that mindset to click with our product design, A lot better. targeting those founders and usually what they are looking for is fundraising, go to market product building, as well as a database which remembers everything and can retrieve the right information and for them, we basically serve all the purposes that they are looking for.

this is an co-founder who can do fundraising, and can remember everything. never forgets I would say this positioning usually is quite attractive to people. but then the next step is we have to explain, okay, what your AI co-founder can do for you.

What does it mean? Yeah, exactly. Yeah. it's challenging when you go to market, for every new category, it's hard to find a very specific word or [00:38:00] very specific,value proposition that you can describe with accuracy, but also simplicity.

exactly. is that part of your story that clicks with people the fastest,AI co-founder,

Yeah, I think so. first of all, it's catchy. A lot of people, all you are curious about what AI co-founders can do. But that's a basically a signal that it catches people's attention first.

And so we try to build all these capabilities that a real co-founder can do. we recently launched our fundraising agent, and it can help the founders at early stage get access to investors by, we making a warm intro. it can help founders. Practice their pitch by having an AI investor to coach them.

the mock pitch. So starting from fundraising, and the next step that we are building is go to market agent and then product agent. So basically this will guide a founder, to get things done from step one all the way to, getting the company founded and having the real teammates

Onboarded.

you guys sound like you [00:39:00] can work with a lot of VCs, like helping their portfolio companies to help them incubate the companies like the founders

as well, right? indeed. We're, starting to partner with some VCs, e especially early stage incubators.

do you feel you're spending more time trying to convince people that your product is different or you're trying to convince them, that they have this problem? Or that the problem matters.

The problem exists. So whenever I tell people, Hey, context loss is one of the biggest problem, they're like, yeah, exactly.

Now, the next thing is to tell them why we can solve this problem. so I think every time when people are people are pitched with this idea about, we solve the context loss for teams by building this AI memory for teams which never forgets.

people always think that this is something that they need.

And we actually spent more time, doing demonstrations to people,and showing them that how this can be done. [00:40:00] how do they do queries?

How the queries can be accurately done? And when people see the results, they're like, wow, this is very interesting. I wanna get on board.

Do you feel you're in a very crowded space?

Yes and no. When you're talking about AI SaaS tools or AI productivity tools, there are a lot. You can see like hundreds of tools coming out every single day.

and as founders, I get anxious too sometimes with seeing those tools. it's not only about seeing the competition, It's more about, am I catching up with the trend? Because there are always like so many things coming up every day. But no, because all these tools provide one feature or another. There are more like. your hands and feet. it's less about your second brain that can actually remember everything. So when it comes to having a knowledge base, which can remember everything for the team, I don't think there are a lot of people, a lot of companies that have done it, yet.

and Glen does it well, but it's for enterprise, and it's not self-serve. So for startups, especially early stage founders, there [00:41:00] aren't many solutions yet.

I don't know if Glean goes into the depths that you guys are going into though.

Yeah, exactly right. yeah, like we know only have the right information, but also does execution,for the teams and it's like everything in one.

so we're also very different from Glee. and that's why I'm saying like, we don't actually see direct straight competition, but when it comes to AI productivity, there are a lot of tools out there.

That's true. Yeah. So in the Yes. Part of your answer, when you feel like, okay, it is somewhat, crowded space, what's been the hardest part of trying to stand out in that space?

Awareness actually, because we're a startup and it's very hard to, make everyone aware of this tool.

It's like this problem exists in teams. For long time already, and people are looking for solutions

It also, I would say from the product standpoint, there's an interesting loop. you have to find the value of this product for you [00:42:00] to invite team members.

Onto this product to chat with the team members. But the problem is, if you don't invite the team members to this product and you don't have a lot of memories with this product, you wouldn't feel the differentiation that much.

The more you use this tool, the more useful it is for you. And so this like code start problem is also something that we need, improve on.

Yeah. We all have that problem. are you concerned if, let's say companies like Notion or Slack start building long-term memory tomorrow?

I think they have to memory to AI is Data to internet in the internet era, it's impossible that you have, you built something without data. so it's also impossible that you build an AI tool without giving AI the right context and still relying on human beings to provide AI with the right context.

So they will do, but this whole thing is that it takes a lot of effort to actually make this AI memory work. and we have done a lot of work,previously. so basically it's not only about making the [00:43:00] right framework work, but also about. Making the infrastructure right. basically how do you structure your previous data, and how do you structure your whole team to build this whole thing.

And for bigger companies, I would say, they tend to move a bit slower. And that's exactly the advantages for startups. They move fast. we are fearless and so I,I'd say that we designed the product as an AI native product from day one. It's not only about building the memory and having an a AI copilot that works.

Your sales tool. It's about thinking about how AI can help people achieve better productivity from first principles perspective, like making AI a true team member. So the things that we talked about, having AI in the loop and having AI following up your tests from the beginning to the very end.

Those are designed within our system from day one, and it's very hard for a more traditional SaaS company to rewind everything that they have done just to build a new AI native tool from [00:44:00] day one.

Yeah. It's not gonna happen overnight. I believe that. No way. You have years of research in like long term memory.

It's hard.

Yeah, exactly. And I've personally worked at Meta by Dance ai and I understand how Innovators dilemma is a real thing. Yeah.

how do you see Tener in three years? I wanna know what's what to expect, what Tener.

Yeah. we definitely will redefine how human and AI collaboration looks like. you can see a lot of hints in our product now already. I think about the previous SaaS generation as something like Lego, right? You have Lego toys, you have different modules you can put together.

Those modules are standardized, so you can put everything together as you want. But I think about ai, a native product as 3D printing. So when you have 3D printing, you actually don't need Lego toys anymore. You just print everything on demand personalized for you. So you can see tanker tool as the basis [00:45:00] of a 3D printing.

You can build everything on top of that. the premise is that this 3D printing machine knows everything that you want. It has the memory. Once we have the memory, you can build a lot of things on top of that. The next step that we are gonna build is better workflow. Workflow nowadays is.

Fragmented in different tools and in different people's minds. so it's called Rigid Workflow, but we are trying to transform the workflow into flow. So when you just talk to the teammates, talk to your AI and AI builds things for you. You don't have to worry about switching between tools.

You just have your ideas flow and talking and that's called the flow.

And the next step is having better collaboration. How do we nowadays people collaborate in chat, then they have to move the context to, notion docs, to whatever documents that they use.

on top of the, the good memory. when these things happen, then we can truly design an AI native way of collaboration for companies. And we can truly enable, [00:46:00] smaller teams in the AI era, especially the solo entrepreneurs or the smaller teams with super entities, to collaborate together.

I love your, analogy about like 3D printing versus Lego, so no longer piecing things together, but things can just be hyper-personalized essentially. Whatever you want. It's possible.

I would love to have just my mind just go into the flow and things can just happen.

Yeah, exactly. That'll be awesome, right? Yeah, totally.

Thank you so much. Yeah, really nice chatting with you today.

I learned a lot about, the architecture and your product design, philosophy and what kind of pain points and problem you guys are solving. And I do think this is extremely valuable and I think the technology is here and people are waiting.

I'm looking forward to your updated launch soon, right? Yes. end of September. okay. Yeah. Do you think you can share, like a beta tester link or whatever, discount link you can share with my audience and we can [00:47:00] let them try it out and give you feedback.

Yeah, definitely. We'll share with you a discount link, very soon.

Thank you so much. Really nice meeting you.

Thanks a lot.