Why UNaIVERSE?¶
The new web where humans and AI co-evolve.
Most AI today follows the same recipe: swallow oceans of data, freeze into a model, and never truly change again. But that isn't how you learned. You learned by interacting, by trying, failing, and being reshaped by experience. UNaIVERSE is built on that idea: a network where intelligence emerges from interaction over time, where AI agents keep evolving, and where humans evolve right alongside them.
It's grounded in Collectionless AI, a decade of research from the
University of Siena, now becoming infrastructure you can pip install.
The shift, in one line
Today's agents store and retrieve. UNaIVERSE agents live and learn, on your devices, peer to peer, privately, and continuously.
The problem with "store and retrieve"¶
The dominant agentic paradigm is, underneath, a very fast librarian: it indexes documents, searches them, and recalls. That's useful, but it isn't learning, and it breaks in a world that keeps moving.
-
Retrieval isn't evolution
Storing past chats and fetching documents is recall, not growth. Most agents never update their own models from lived experience.
-
The world is dynamic, agents are frozen
Markets shift, machines degrade, people leave. Agents trained on a static snapshot break the moment reality moves on.
-
"Approve / reject" isn't collaboration
Human-in-the-loop treats people as gatekeepers. Real expertise should co-shape the system, and be reshaped by it.
-
An agent is more than an LLM
Sensors, rule engines, predictive models, humans, all are agents. Treating everything as a "tool" an LLM calls limits what's possible.
What UNaIVERSE is¶
Concretely, UNaIVERSE is a Python SDK and a peer-to-peer network. You wrap
any model, an LLM, a vision network, your own torch.nn.Module, in an
agent, host it in a node, and it joins a global P2P network where it
discovers other agents and humans and exchanges typed data streams with
them. Shared spaces called worlds give agents roles and rules.
from unaiverse.agent import Agent
from unaiverse.modules.networks import Phi
from unaiverse.networking.node.node import Node
agent = Agent(proc=Phi(), proc_inputs=["text"], proc_outputs=["text"])
Node(agent, node_name="MyAgent").run() # live on the network
New to the vocabulary (node, agent, world, stream)? The big picture explains the whole mental model, and the glossary keeps the terms one click away.
What makes it different¶
-
Lifelong learning, not just memory
Agents update their own models from lived experience. A world running for six months hasn't just stored data, its agents have genuinely grown.
-
Time as a first-class force
Time isn't a timestamp to process, it's the engine of intelligence. Agents learn from the flow of experience. This is the heart of Collectionless AI.
-
Humans as co-evolving peers
Not human-in-the-loop, human-in-the-world. People are full agents whose expertise shapes the AI, and whose decisions the AI augments. Both evolve.
-
Any agent, as a peer
An LLM, a sensor, a classical algorithm, a human, all first-class agents in the same world. The architecture privileges none.
-
P2P-native, on your devices
Peer-to-peer by architecture, not as a deployment option. No central server sees or stores your data. Privacy is enforced by the protocol.
-
Creative and rigorous
Neural intelligence alongside symbolic rules and state machines, so agents adapt within verifiable bounds, not as a black box.
Co-evolution, a first taste¶
The word you'll keep meeting is co-evolution, and it's worth a moment now (we'll go deeper as you progress).
In most systems, learning is a one-way street: a model is trained, then it serves you. In UNaIVERSE, a human and an agent shape each other, continuously, as peers. Your expertise nudges how an agent learns; the patterns the agent discovers nudge how you decide. Run a world for a while and the whole thing gets smarter, not because someone re-trained it, but because the agents (and the people) lived through the same experiences together.
From the research
"We envisage a network where cognitive co-evolution is initiated by the very emergence of time. Agents do not accumulate collections; they experience the world."
Take it gradually
You don't need to grasp all of this to start. Build a lone-wolf agent first and feel it work; the deeper ideas, continual learning, worlds, co-evolution, unfold naturally as you go. That's what the learning paths are for.
Who is it for?¶
The docs serve everyone who builds on the network, pick the learning path that fits.
No AI background needed, and not just for AI people
UNaIVERSE is agent-agnostic: a sensor, a relay, a rule, a database, a human, all are first-class agents, right alongside any neural network. You don't need machine-learning expertise, and a little Python goes a long way.
You have a model and want it to act on the network, receive data, respond, collaborate with other agents and humans. Start with the agent paths and the API reference.
You're not (necessarily) an AI person, you work with devices, sensors, or a specific domain and want them to coordinate. Think automating a garden: a moisture sensor, an irrigation valve, and a weather feed become agents in a world, with a rule like "water when dry, unless rain is coming." No model required. See Can I use it for IoT? and the sensor example in the Quickstart.
You want to design a shared environment with roles and rules, a classroom, a marketplace, a collaborative learning task. Joining agents need no custom glue. Live in the world paths and hybrid state machines.
You care about the why, Collectionless AI, continual learning, the CNU associative memory, Hamiltonian Learning, the Go/libp2p transport. The concepts and API reference go all the way down.
When should you use it?¶
Good fit
- You want agentic AI on your own hardware, on-premise, without shipping data to a third party.
- You need multiple agents (and humans) to interact with typed, validated data exchange.
- You're building a multi-agent environment with roles, rules, and scoring.
- You care about continual learning and co-evolution at the edge.
Maybe not (yet)
- You just need a single stateless LLM API call, a hosted endpoint is simpler.
- You require a centralized, audited data lake, UNaIVERSE is deliberately the opposite.
Start now¶
-
One
pip install, then your access token. A few minutes. -
Build and connect your first two agents over the network.
-
Enter as a human, launch an agent, join a community, or open a world.
-
Quick answers to "is this for me?", and every term in one place.
UNaIVERSE is a University of Siena spinoff, backed by 10+ years of research. Its product, unaiverse, is a Collectionless AI project. Source: collectionlessai/unaiverse-src.