Typically, neural networks operate in a «here-and-now» mode. They provide answers based on what they observe at the current moment. But if you end a conversation and return a week later, the context has vanished. You can, of course, save chat history, but that is not the same: such memory is «flat»; it does not help distinguish what was important and what was just «random background noise».
Nvidia approached the problem differently. The company introduced a technology called Memory for the Blockchain of Life — abbreviated as MBL. The idea is to give AI agents the ability to remember events and details the way people do: not everything indiscriminately, but what mattered — what is connected to emotions, actions, and people.
What MBL is and why it is needed 🧠
MBL is a memory system for agents operating in virtual worlds. It is built on the open-source MemoryForge framework and integrated with the Nvidia ACE platform. In short, it is a set of tools that allows digital characters to recall the past, link events, and behave more naturally.
Why is this necessary? Imagine a game or a virtual reality environment where you interact with non-player characters (NPCs). Usually, they either repeat the same phrases or act strictly according to a script. With MBL, a character can remember that you have met before, what you said to them, and how you acted last time. This makes communication livelier and more meaningful.
According to the developers, such memory will be useful not only for games. It can be applied in training simulations, digital assistants, virtual consultants — anywhere personalization and continuity of interaction are important.
How this memory is organized
At the core of MBL lies the MemoryForge framework, which Nvidia released as an open-source project. It works on top of large language models and adds a structured memory storage system.
Memory here is divided into several types:
- Episodic — specific events: «we met in the park», «he asked for help».
- Semantic — general knowledge and facts: «this person is a doctor», «the city is located in the north».
- Procedural — skills and action patterns: «how to open a door», «how to react to a request».
These types of memory interact with each other. An agent can recall not only what happened but also how it relates to other events, what conclusions can be drawn, and how to proceed.
A key feature of the system is that it uses the concept of «importance». Not all events are saved equally. If something was emotionally significant or influenced the agent's actions, it will be remembered better. If it was just background conversation, it may be forgotten — much like human memory.
Demo: Inworld and Baobab Studios showed how it works
Nvidia demonstrated the capabilities of MBL in two projects.
The first is from Inworld Studio. It is an interactive virtual reality experience titled The Last Bounty Hunter. You play as a new character in a Wild West town, surrounded by non-player characters (NPCs) with memory. They remember how you behaved, what you said, and what decisions you made. Their attitude toward you changes depending on that.
Technically, it works like this: agents use models based on Nvidia NIM and the RTX AI Toolkit. Memory is updated in real time; characters can refer to past dialogues and actions. Developers say this approach enables deeper plots and enhances the sense of a «living world».
The second project is from Baobab Studios, called Sonderlust. This is an animated story in virtual reality about a character named Sonder. Here, memory also plays a key role: the protagonist experiences a series of events that gradually shape his personality. Memory allows him to grow, change, and react to the environment not mechanically, but by taking past experience into account.
Why this is harder than simply saving logs 📝
At first glance, it might seem obvious: just save everything to a database and retrieve it later. But the problem is that «everything» is too much, and the vast majority of it does not matter for future interactions.
Human memory works differently. We do not remember every second of our days, but we remember key events well. We link memories — one leads to another. We forget details but preserve the essence. And this is not a bug; it is a feature that helps us navigate the world without «drowning in information».
MBL attempts to replicate this mechanism. The system filters information, evaluates its importance, and links events together. When needed, it retrieves not just a fact but the context. This makes an agent's behavior more natural.
Where this might be useful
Games are the most obvious application. NPCs with memory can make worlds more convincing. Instead of starting a dialogue from scratch every time, a character can continue a topic, recall your previous meeting, and react to your actions.
Virtual reality training is another area. If you are training on a simulator, a virtual instructor with memory can adapt to your mistakes, remind you of past lessons, and give hints based on where you have already encountered a similar problem.
Digital assistants are also a logical application. An agent that remembers your preferences, past requests, and context can be far more useful than a generic assistant without memory.
But there are nuances. The more memory there is, the harder it is to manage. Issues of storage, privacy, and data updates need to be addressed. Furthermore, if an agent remembers too much, it might seem strange or even creepy. Balance is important.
Open source and accessibility
Good news: MemoryForge is available as an open-source project. This means developers can take it, adapt it to their tasks, and experiment.
Nvidia has also integrated MBL into its ACE platform — a set of tools for creating interactive AI characters. It includes support for speech, animation, and line generation. Adding memory completes the picture: now a character not only speaks and moves but also remembers.
Developers can run all of this on local machines with RTX graphics cards or via cloud interfaces (APIs). This makes the technology accessible to both large studios and indie teams.
What's next
MBL is a step toward agents that behave more like living interlocutors. They do not just react to a current request but take the past into account. They can change their behavior, learn from experience, and build relationships.
Of course, a full-fledged «digital personality» is still a long way off. But the direction is promising. The more natural interaction with an agent becomes, the more use cases open up.
So far, MBL has been demonstrated using games and virtual reality as examples. But the principle is universal: any situation where an agent needs to interact with a user repeatedly can potentially benefit from such a memory system.
It will be interesting to see how developers begin to use this tool and how users react to agents that actually remember.