Our system doesn't need a model to do the retrieval
· Richard Wollyce
What's up everyone. I need to take this moment to talk about the project I'm building, Ulpia. And this is also going to be the first video on the channel. This channel is going to be focused on talking about technology and other things I find interesting, but mostly we're going to talk about software development and technology.
This one is going to be quick, but I hope it's informative about Ulpia, and also about the reason why I decided to build it. So let's go.
I've noticed that artificial intelligence really is part of our lives now, especially for anyone who works in software development. It's something we can't escape anymore, especially for anyone already working at a production level, anyone already shipping product to the market and so on. These days there's no way for us to do things purely by hand, coding everything by hand. These days you need the efficiency that AI helps you get. And of course, following software engineering fundamentals so you don't ship anything without quality.
Documentation is one of the main pillars
And one of the things that is obvious, that improves the output of your AI agents a lot, of your work with artificial intelligence, is documentation. The better documented your system is, the better any maintenance you do there is going to be, or any implementation of new features. And that's why I think documentation is one of the main pillars today for anyone who develops software.
And thinking about that, we get to another point, a point that is correlated to it, that doesn't necessarily only work for people who develop software. It solves problems for other people, in other areas, but in particular it solves them for anyone who needs an agent to have a specific, detailed and persistent memory.
We get to this AI Memory problem
So we get to this AI Memory problem. And it was thinking about this problem that I identified some possible solutions for what I was facing in my day to day.
I know some solutions exist. There are solutions like Mem0, there's Zep, there's Letta, there are, well, other startups. Both open source and closed source too, trying to solve this AI memory problem. But the Ulpia project, I believe it's going to be even more revolutionary, let's put it that way, because I decided to build it focused not only on AI Memory, but also focused on people who want to have things locally.
The privacy of that data
Because that's where another point comes in. Beyond the quality of the documentation and of the records we need to keep when we're developing software, there's also the privacy of it. The privacy of that data, the way you store that sensitive information, business information and so on. These things need special care, especially when you're dealing with things that are a little more delicate.
And that's why I believe the future of memory for artificial intelligence shouldn't be a memory hosted in a cloud service, but a memory you can have locally, and where you can run both cloud agent models and local agent models.
Of course, today that's still not a reality, as a matter of infrastructure. Hardware is still very expensive these days. So having laptops and computers with enough RAM, enough processors, to be able to run artificial intelligence locally, still isn't very appealing, I know that. But I believe very soon it will be.
And that's why I already decided to take the initiative to develop this system, which is going to give you the possibility of having a local artificial intelligence memory, with expressive functionality and very low latency, because we use Rust to do the retrieval path for the information in those memories.
And in this case, I built Ulpia precisely so that you can do the retrieval and the storage of those memories in a way where you don't depend on an artificial intelligence model administering every action. That's one of the main points of Ulpia.
The other solutions that exist on the market
The other solutions that exist on the market, they use the model to record those memories. We can use it too. But they also use the model to do the retrieval, to read that information. And our system doesn't need a model to do the retrieval.
We use an intelligent system for capturing information in your memory base, one that ranks those memories in order to rank the files that are compatible with your input, with what you're sending for your artificial intelligence to deal with. And only after we have a ranking do we actually hand it to an AI model, whether local or cloud, well, an external provider. Then that model starts to act on the information that file brings, that is, that memory file.
This repository is going to be open source
But anyway, that's it. Today I'm not going to be able to go into much detail, but I needed to present the project. I'm going to leave the repository link, because this repository is going to be open source, it's going to be public. So you can contribute to the evolution of this project, and you can also self-host it, you can use it on your own machine and integrate it with any kind of LLM you want. Whether local or cloud, whether, for example, from an AI provider like Anthropic, OpenAI, DeepSeek, well, any LLM model you want to use.
You can implement this system. It works over MCP locally too. So you're going to have the system, and your agent is going to make calls over MCP and it's going to receive that information over MCP too, to talk to the system.
And that way you're going to be able to save on reading tokens, because you're not going to have to wait for your LLM model to read the file to find which file has the subject you're looking for.
And you're going to be able to, obviously, like I said, handle the memory on your own. You yourself can write information into the memory files, and you can also send an agent, a model, a specific agent that you might have built as well. That's another thing that's going to be left for another video, which is the format we use for building specialised agents. But anyway, you can also set that model to record in memory whatever you need, the information you require.
And that's it, folks. I hope you like the project. And we're open to feedback, obviously, so you can leave comments, or open issues over on the GitHub repository, which we'll be taking care of, and we're going to improve this, so it fulfils this purpose, which is to help make artificial intelligence even more efficient, even more secure and with even more privacy.
And that's it, thanks.