Skip to content
YC Root AccessYC Root Access

This Startup Is Trying To Solve The AI Memory Problem

While LLMs continue to evolve, they still struggle with memory. The startup Mem0 is working to change that by building the memory layer for AI agents. In this episode of Founder Firesides, YC’s Nicolas Dessaigne sat down with co-founders Taranjeet Singh and Deshraj Yadav to discuss why agents need persistent memory to improve over time, how Mem0 reduces cost and latency compared to native context stuffing, and why memory must remain neutral across models as AI becomes more agent-driven. Chapters: 00:05 What Is Mem0? 00:49 Traction & Open Source Adoption 01:24 Why Memory Improves AI Agents 02:01 Saving Cost and Latency 02:31 Founder Origins & YC Pivot 05:13 How Mem0 Works Under the Hood 06:04 Hybrid Memory Architecture 07:10 Custom Memory Rules & Expectations 08:00 Real-World Use Cases 10:05 Competing With Model-Native Memory 11:48 Fundraising & What’s Next

Nicolas DessaignehostDeshraj YadavguestTaranjeet Singhguest
Jan 23, 202617mWatch on YouTube ↗

EVERY SPOKEN WORD

  1. 0:00 – 0:05

    Intro

    1. ND

      [on-hold music]

  2. 0:05 – 0:49

    What Is Mem0?

    1. ND

      Today I'm joined by Taranjeet and Deshraj, the founders of Mem0. They just announced a 24 million raise to be the memory layer for AI. Congrats guys, and welcome.

    2. DY

      Thank you.

    3. TS

      Thank you. Thank you for having us.

    4. ND

      All right. Tell us what is Mem0 today?

    5. TS

      First of all, thank you for having us. Mem0 is building memory layer for AI agents.

    6. ND

      Mm-hmm.

    7. TS

      Right now everybody is trying to create an AI agent, and all of them are using LLM, but there is a fundamental issue. LLMs are stateless. They don't remember things like human remembers, so we are trying to fix that for every agent and every AI app that anybody's creating out there.

    8. ND

      It means that e- every time someone, uh, does a prompt, then the agent starts from scratch-

    9. TS

      Yeah

    10. ND

      ... and not memo- don't remember whatever happened before.

    11. TS

      Yeah, yeah, yeah.

    12. ND

      What's the current

  3. 0:49 – 1:24

    Traction & Open Source Adoption

    1. ND

      state of the business?

    2. TS

      The current state of the business is doing well. We are the most adopted solution in the market in terms of traction. We recently crossed 14 million Python package downloads, 41,000 GitHub stars.

    3. ND

      Yeah, you're open source, right?

    4. TS

      Yeah, we are open source, and we are used by thousand of companies across. We not only power memory for companies, we also s- provide memory for major agentic frameworks like AWS, Agents SDK, CrewAI, Flowise, et cetera. And recently we announced our $24 million funding round.

    5. ND

      And what are the main benefits here? Of c- of course, agents get that memory, but what is the benefit for them?

    6. TS

      Uh,

  4. 1:24 – 2:01

    Why Memory Improves AI Agents

    1. TS

      the main benefit for anyone who is using a solution like Mem0 in their AI agent is, like, their agent improves, uh, over time. It's like, let's say if I'm building a, a trip planner agent, and I'm like, "Whenever I'm going to New York, I want to stay near Wendy's," the agent will remember this preference, and the next time whenever I'm automatically booking a trip to New York, it will just surface it.

    2. ND

      Mm.

    3. TS

      So that's the first benefit wherein, like, the new AI apps and agents that are coming, they know you in the best possible way. And then apart from that, for developers, we also help them save cost because we optimize their prompt, and we also help them save latency.

    4. ND

      So how, how do you save cost? That's interesting-

    5. TS

      So the-

    6. ND

      ... 'cause you

  5. 2:01 – 2:31

    Saving Cost and Latency

    1. ND

      have more data.

    2. TS

      No. So the most naive way wherein, like, anybody who wants to have some sort of memory layer is by passing everything into the context window.

    3. ND

      Mm.

    4. TS

      When you pass everything into the context window, you are, you know, sending more tokens, so more cost and more latency. We on the other hand will optimize for the right and the most accurate information. That's how we save cost and tokens.

    5. ND

      So instead of like, uh, throwing everything that happened and then copy-pasting that in the prompt-

    6. TS

      Yeah

    7. ND

      ... you can be more smart about what you share.

    8. TS

      Yep, yep.

    9. ND

      Awesome. Tell us more about how you started. And-

    10. DY

      [chuckles]

    11. ND

      Where are you

  6. 2:31 – 5:13

    Founder Origins & YC Pivot

    1. ND

      from, guys?

    2. TS

      I'll go first. You know, I'm from India. I moved to US last year. And, uh, my journey involves me taking seven attempts in starting a company.

    3. ND

      Mm-hmm.

    4. TS

      I took the first attempt in 2012. Uh, this is the seventh attempt in me starting a company.

    5. ND

      [laughs]

    6. TS

      And [laughs] I got lucky that I got Deshraj as a co-founder and CTO in this journey along. And instead of-

    7. ND

      When did you meet him?

    8. DY

      We met actually, uh, during our undergrad. Uh, we are undergrad friends. We know s- each other since like 2012, so been 13 years. We did a lot of projects, like, during our undergrad, uh, a lot of side projects, and always wanted to, like, start something of our own. TJ was doing a lot of things in India. I was here working at Tesla Autopilot, uh, where I was leading the AI platform team. Um, and then he was shifting here, and we were like, "Let's build something together."

    9. ND

      See-

    10. DY

      And that's how, like, this whole story came together.

    11. ND

      And then you joined forces and applied to YC.

    12. DY

      Exactly.

    13. TS

      Yeah.

    14. DY

      Yeah.

    15. TS

      Yeah.

    16. ND

      And I remember when you applied, you actually joined us with another idea, not Mem0, right? What happened? Can you tell us a little bit more of that story?

    17. TS

      Yeah, so, so we got into YC in Summer '24 batch, and we were interviewing with you around March and April of 2024. At that time, we were going through a shift from not just focusing on RAG and focusing more on solving the statelessness problem of LLMs. When we were applying, the company name and the product name was still Embed Chain, but we were, uh, behind the whole tinkering with the memory idea. Uh, the genesis of this idea is from a app that Deshraj and I built in December 2023. It was called Sadhguru AI. Sadhguru is a very famous Indian yogi.

    18. ND

      Okay.

    19. TS

      Uh, that app went viral in India, and one of the common feedback from that application was that this app is cool, but it is kinda dumb because it doesn't remember anything about your meditative journey.

    20. ND

      Yeah.

    21. TS

      So that's when we started thinking like, "Why is this happening?" And we realized that it's because LLMs are stateless. RAG is one way to give context about a knowledge base. But then Deshraj and I kept thinking about it, and we were like, "Th- this is something deeper than that." And that's how we started tinkering with, like, early versions of Mem0 and Embed Chain. And we remember, uh, when both of us met you for the first time at YC Retreat, and, uh, we were, like, talking about all the algorithms that we have developed, and you were like, you know, "Why haven't you launched yet?" And within, like, 36 hours, we ended up launching from Embed Chain to Mem0, and since then the journey has been really great.

    22. ND

      [laughs] That's awesome.

    23. TS

      Yeah.

    24. ND

      That's awesome. Yeah, like the, the traction, the open source traction has been very, very impressive.

    25. TS

      Yeah.

    26. ND

      People started adopting right away, right? It's, uh, it's so, uh, so funny to see that the idea came from the side project that was a consumer app that took off pretty much.

    27. DY

      Right.

    28. TS

      Yeah.

    29. DY

      Yeah.

    30. ND

      That's, that's excellent. Let's go back to the product. So how does that work exactly? Like the inside of the product. I'm a developer. I implement you in my product. How is that going to work?

  7. 5:13 – 6:04

    How Mem0 Works Under the Hood

    1. TS

      I mean, I'll just cover from, like, a high-level perspective. So the product is composed of two primitives. One is, like, adding a, adding a memory, and the other one is, like, searching a memory, and these primitives reflect in our APIs. Uh, so we make it very easy for a developer to give us any, you know, sort of data they feel is important on a user level. We try our best to understand what is important from that, and then we build a state on top of it. So let's say if you're, like, building, like, a personal companion, you'll give it, give us your chat responses. We'll try to extract meaningful information from it, and we'll try to build a state on top of it. Building a state is important because that helps us in understanding how the evolution is happening. And then whenever, like, the user is coming, let's say the user is trying to, you know, start a new conversation, we'll give you the most important conf- you know, information out there. Behind the, uh, hood, we go through, like, a technical process which Deshraj

  8. 6:04 – 7:10

    Hybrid Memory Architecture

    1. TS

      can cover.

    2. DY

      Yeah. Yeah. So, um, our algorithm is actually like a hybrid data store architecture that we have implemented where we are actually, uh, whenever some unstructured information that comes into Mem0, we basically try to, like, uh, classify that into, like, a key value pair depending on what kind of information it is, or as a semantic chunk, or as, like, a graph memory which we call basically where you are trying to basically-Create like relations between like different facts that you're collecting so that later whenever developer is actually requesting like, "Hey, uh, give me relevant information that you know about me," we go and like pull information from all these three sources.

    3. ND

      Okay.

    4. DY

      And we do it effectively and do it in real time, so that helps you basically have like a really low, uh, latency in terms of retrieval and still have like high accuracy. And by doing all of this, you are also like able to like save cost and latency because, you know, instead of putting everything into the context window, which is gonna be expensive and, uh, like slow, so it just, uh, solves the problem for them. Yeah.

    5. ND

      To be able to create all of that presentation of the kind of the memory of the end user, do you need to understand the, the actual application? Do you know ... Do you need to understand what's the actual purpose of the application?

  9. 7:10 – 8:00

    Custom Memory Rules & Expectations

    1. TS

      That's the part which is like somewhat hard, but it's also easy, where like we put in like a lot of effort to understand what the developer is building. So during onboarding of the product, we try to understand like what are you trying to build, and memory is an expectation problem. So like for you, uh, your memories might be different for me, my memories might be different even for the same task, right?

    2. ND

      Right.

    3. TS

      So we try our best both on the open source, on the, on the cloud version, that you can customize anything that's possible.

    4. ND

      Mm.

    5. TS

      So you can come down and you can be like, "Hey, I don't want like this kind of memories to be captured. I want this," and all of this is in plain language, plain like natural-

    6. ND

      Plain language.

    7. TS

      Yes. So you can do that.

    8. ND

      It's going to be interpreted by an LLM.

    9. TS

      Yeah. And then we like try to form rules on top of it, and then we read in the pipeline, we, we update all the memories for you.

    10. ND

      Mm. Do you have any, um, uh, good examples of use cases, like things that people have done

  10. 8:00 – 10:05

    Real-World Use Cases

    1. ND

      with the product?

    2. TS

      When we talk about memory, I think like memory should be a default primitive whenever you're building an AI application. A couple of high-level use cases that we have seen across, uh, people try to use some sort of memory solution for efficiently managing their context and coding agents.

    3. ND

      Mm-hmm.

    4. TS

      People try to use some sort of memory solution when they are building a personal companion. People try to use, you know, memory solutions in education wherein they want to remember the learning trajectory, in healthcare wherein they want to remember everything about the patient and the medicine. In finance, they want to remember the entire trade history. So it's like wherever you're building, you know, like an LLM-based application-

    5. ND

      Yeah

    6. TS

      ... and you want it to get better over time-

    7. ND

      Mm

    8. TS

      ... you should need memory and you should use Mem0.

    9. DY

      Yeah. We have also like started seeing this very interesting pattern now where instead of capturing memories about humans, people are actually now building more and more agents, so they want memories about the agent.

    10. ND

      Yeah.

    11. DY

      So they want to capture more and more of that as well.

    12. ND

      Yeah. Excellent. Is there sometimes an issue when, uh, memories become stale? Kind of like is there any kind of decay of the memories-

    13. TS

      Right

    14. ND

      ... that you want to use them differently?

    15. DY

      Yeah, so that's a very good question actually. So we have been seeing this pattern from our users where different developers actually ask for like different kind of decay. Uh, sometimes like customers are like, "Hey, I want like a hard decay where after six months I don't care about any memory." Sometimes developers ask for exponential decay where they care more about the recent stuff, but they want to forget as memories get super old. And sometimes we have also seen developers ask about like certain things depending on their application. Let's say someone is building like a travel planner agent, uh, things that are related to travel preferences always matter no matter how old they are. So, but they still want to like forget other stuff which are probably not that relevant. So we are seeing these interesting patterns, and we have like implementations of each of these decay mechanisms.

    16. ND

      Is that same thing like they are going to describe what they want like in plain language?

    17. DY

      It is kind of, but there are certain other knobs that you can also tune basically to enable that.

    18. ND

      Okay. Excellent. Uh, let's take a, a step back quickly. OpenAI recently launched this memory layer for OpenAI. Like other labs are building the same of course.

    19. TS

      Mm-hmm.

    20. ND

      Like are you still relevant in that world? Like what does it mean for

  11. 10:05 – 11:48

    Competing With Model-Native Memory

    1. ND

      you?

    2. TS

      Yeah, I think it's a good thing for us wherein like all the big labs are launching memory, and the memory is available in their consumer app, uh, offerings, and it's a matter of time that it becomes available as an API for developers. Uh, the fact that it's good for us is because they're educating the market that you need memory as a default primitive in any AI application. But for us it's good because developers are using multiple LLMs whenever they're building an AI application.

    3. ND

      Mm.

    4. TS

      Right? And memory is not just read only, memory is write only also. So in that case, as in, you know, best engineering practice and even like as a first principle thinking, you would not want to tie your memory to any, you know, model provider out there. For model provider, i- memory is the next mode because models are becoming a commodity. But for a developer, because they are using multiple LLMs, it should be decoupled.

    5. ND

      It's able to own the memory-

    6. TS

      Yeah

    7. ND

      ... and not, like still always have the option to change a model.

    8. TS

      Yeah.

    9. ND

      Okay. That makes complete sense. Uh, and, uh, and on that note, you, you mentioned that you work, uh, with other kind of frameworks, other partners like AWS.

    10. TS

      Mm-hmm.

    11. ND

      Like how, how does that work? Like, uh, people using these frameworks have you out of the box?

    12. TS

      Yeah. So like, uh, we not only serve the customers, we also provide memory to agentic frameworks. Like we are the exclusive memory provider in the A- AWS agent SDK called Strands, and then we power memory for all the other major agentic frameworks like CrewAI, Flowise. The idea is that memory is something which should be neutral to anything. It should be neutral to framework, it should be neutral to your model provider, it should be neutral to LLM.

    13. ND

      Mm.

    14. TS

      So that's how we give it like as a very simple tool call in any of the agentic frameworks out there, and developers who are using multiple frameworks even can have their memory in a central fashion and like keep it like decoupled from anything

  12. 11:48 – 17:21

    Fundraising & What’s Next

    1. TS

      out.

    2. ND

      You recently announced your like kind of big fundraising, uh, like 24 million from a Basis Set and Peak XV and others. Like-

    3. TS

      Mm-hmm

    4. ND

      ... uh, why did they see in you? Like what convinced them to invest?

    5. TS

      First of all, it's a, you know, like it's a 24 million seed plus Series A round. Our seed was done by, led by Kindred. Series A was led by Basis Set. Uh, Basis Set participated in our seed round, and we have known like Lan, uh, very well for over a year. And, uh, we really like the partnership so far. And, uh, when we decided to raise, uh, they quickly got back to us with like the fact that they wanted to double down on us, and we, you know, were fortunate enough to raise from them, Peak XV, Kindred, YC, and a lot of other angels as, like great angels as well.

    6. ND

      That's awesome. So they saw like the traction, pretty much they were already insiders. They knew things were working-

    7. TS

      Yeah

    8. ND

      ... and wanted to double down.What are you going to do with all this money?

    9. TS

      I mean, we want to build the best possible memory product on the planet. Uh, we are u- going to use this fund to hire the best possible team so that we can [laughs] build the best possible product. Nothing else.

    10. ND

      How big is the team today?

    11. TS

      Uh, so we are, like, now 10 people, uh, split across India and SF office.

    12. ND

      Mm-hmm. And so what role are you hiring for?

    13. TS

      So we are hiring across. We are hiring for, uh, you know, applied AI. We are hiring for full-stack. We are hiring for forward deployed. We are hiring for, uh, GDM engineer. Uh, Deshraj can give, like, a one-minute pitch also. [laughs]

    14. DY

      No, I think a- across the, across the engi- uh, engineering, uh, organization, basically we are trying to, uh, double down and, like, ship as fast as possible, and that's why, like, full-stack, forward deployed, applied AI, research engineer, um, front-end, back-end, all these things.

    15. ND

      B- both here and in India?

    16. DY

      Both, yes. That's correct.

    17. TS

      Yeah.

    18. ND

      Okay.

    19. TS

      And, like, whatever work we are doing, I think, like, on engineering side, it's, like, one of the most challenging work out there. You have to build, like, a low latency infra product.

    20. ND

      Yeah.

    21. TS

      And you're solving an expectation problem, so you'd have to, have to cater to every user, and you have to make sure that memories work well, reliably, and scale.

    22. ND

      Excellent. And what do you see next happening in that, uh, world of memory for AI? Like, what will be your vision, like, a year or two years from now?

    23. TS

      Yeah. We recently came up with this, uh, you know, three sentences, uh, while brainstorming. We call it, like, make it work, make it neutral, and make it portable. Right now, we are in the first two phases. We're in, like, memory works, but memory should work very well across any domain out there, so we have to keep pushing the frontiers of that. And we want to keep it neutral. It should not be tied to anything out there. But the broader theme that we see, maybe which happens in, like, 5 to 10 years from now is, like, we are going through a technological shift. So far, we as humans have been interacting with any technology using swipe, scroll, and click, but that's going to change. It's going to be a lot of agentic interfaces. For the first time, you and I can chat with any app or can talk with any app, and the app can talk back to you.

    24. ND

      Yeah.

    25. TS

      So we, you, everyone is accumulating rich personal data and that. And history, you know, has shown us multiple times that user expectation always move towards less friction. So five years down the line when you are having, like, a hundreds of AI apps in your life day... on a day-to-day basis, and you would have, like, given every app some custom instruction about how it should be, and you are trying to... and you are trying the 101 app, and you're like, "Why doesn't this app just get me?"

    26. ND

      Mm.

    27. TS

      So that's the future we will be excited-

    28. ND

      It's the portable, like you want the-

    29. TS

      Yeah, yeah.

    30. DY

      Yeah.

Episode duration: 17:22

Install uListen for AI-powered chat & search across the full episode — Get Full Transcript

Transcript of episode Sr1STQP0cds

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome