Why LLMs Need a Collective Memory Layer | Pablo Stern-Plaza

Pablo Stern-Plaza (0:00)

One of the things that's, I think, super interesting in this space is, when you think about these agents, these are not just single agents with single memory. They can have a collective conscience. They basically can share memories across multiple agents, and that makes them incredibly powerful and useful. If one agent solves a problem, it has the best way to solve that problem, and it feeds that information to the rest of the agents, they all get that. And you can only do that by really providing that memory layer in a way that scales, that provides that right level of security and governance, and does so in a way that ultimately helps the outcome of what the agent is doing.

SPEAKER_2 (0:33)

Hello, everyone, and welcome back to MTS. Today, I am joined by Pablo Stern-Plaza. He is the CPO of AI & Engineering at MongoDB. Excited to have a great conversation because today, MongoDB announced a bunch of cool things that you all launched, including the Agent Engine that you just talked about. I know you worked a ton on that product. So can you share a little bit more about that?

Pablo Stern-Plaza (0:55)

Yeah, absolutely. Look, I'm really excited. First of all, I'm really excited to be here, and I'm a product person. So for me, days like today are awesome. As you get technology out into the hands of builders and they start using it and start giving you feedback, that's what this job is all about. So we announced Atlas Agent Engine.

And Atlas Agent Engine is our agentic stack to build, deploy and run agents in production using a memory layer with the right security guardrails and the ability to run anywhere. Yeah.

SPEAKER_2 (1:28)

So I like that you're a product person. And especially when we think about how these agents are actually being used, one of the most important parts about when you're a product person is you have to talk to customers. You have to actually understand what are the big pain points and how they actually want to use these tools in production. When you talk to customers, what are the common themes you're seeing?

Pablo Stern-Plaza (1:47)

Yeah. So look, we talked to hundreds of customers. And I think one of the advantages with Atlas Agent Engine is we had the opportunity to scour the market and we were actually talking to customers that were using some of the solutions that currently exist. And so we started getting a lot of feedback on gaps that they were seeing, that they were looking to solve in another solution. And it came down to three things. And the first one was probably pretty obvious. Because MongoDB is built on JSON, the document model, a lot of our customers were saying, look, it just makes a ton of sense for us that MongoDB, both because you have the document model and because you built all these great retrieval capabilities, like text search, vector search, and you've incorporated voyage and embeddings and re-ranking, the ability to find those needles in a haystack. Because you have all those capabilities, you're just a natural home for memory. We want to use you for that. And in the conversations I've had, I've probably had like 30 customers that are already doing this with us, just using the building blocks that we have. And so for me, that was a bit of a tip of the spear conversation of, well, that makes a lot of sense. What if we gave that to you as a composable component and you could start building on that? And that resonates tremendously with customers because right now a lot of customers are not getting into production. Because when they work on these agents, they're putting all the data into markdown files or wikis or these places where they've realized that that's okay for a prototype. But when you get to production and when you start thinking about that data is becoming your system of record, your core knowledge for the business, you want the same guardrails as all that operational key data. What better place to store that in MongoDB? That was the first thing we heard a lot from customers. Then the second one was, okay, now the other thing that's super important is, this was really true in the enterprise, is security in these agents is top most importance.

Right now, as we're building, there are a lot of building block components, but the ability to drive these security guardrails and do that in a way that's centralized, that not only gives you security from the data layer, but also to role-based access and to be able to isolate workloads. Things that MongoDB already does in Atlas, but extended to agents, was a big gap that customers were seeing. They love the idea that, hey, we could provide that right out of the box.