Distributional Connexed Cognition (DCC): An Experience-Driven Architecture for Artificial Agents
This paper proposes Distributional Connexed Cognition (DCC), an experience-driven AI architecture grounded in the CISR framework and neuro-functional principles, which enables artificial agents to autonomously generate self-referential analogies and accountable decisions through a recursive process of condensing experiential traces rather than relying on pre-programmed datasets.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer
Most artificial intelligence today operates like a student who has memorized a textbook but never stepped outside to see the world. It learns from massive datasets curated by humans, following strict rules written by programmers to recognize patterns and generate answers. While these systems can be incredibly fluent, they lack a personal history; they do not build a library of their own experiences, nor do they adapt their behavior based on what has happened to them specifically. This creates a fundamental gap in how we understand machine intelligence: if a machine cannot learn from its own life, can it truly be said to have its own agency? A new study proposes a different path, moving away from pre-programmed rules and toward an architecture where an artificial agent learns solely by living through its own interactions, accumulating a unique history that drives every future decision.
The researchers, led by Zi An Wang at the University of Ottawa, have developed a system called Distributional Connexed Cognition, or DCC. The core idea is to mimic how the human brain processes information before we are even consciously aware of it. In the human mind, signals from our body and the environment are constantly being received, filtered, and integrated in a pre-conscious "soup" of activity. Only the most important signals rise to the surface to become a conscious thought or a deliberate action. The DCC architecture attempts to replicate this process in a machine. Instead of being told what to do, the artificial agent is given a set of basic needs and a distributed memory system. It receives sensory input, compares it against its own past experiences, and uses that comparison to decide what to do next. If it makes a mistake, that failure is stored as a specific memory trace, influencing future choices in a way that a simple rule-based program could never achieve.
To test this concept, the team created an artificial agent named A1 and placed it in a series of simulated environments, starting with a controlled "nursery" and eventually moving to a more complex, open world. In the nursery, A1 had to learn basic survival skills, such as moving, observing, and eating different types of resources like water and wood. Crucially, the system was not given a list of instructions like "eat water when hungry." Instead, it had to discover these relationships on its own. When A1 tried to eat something that hurt it, like a live metal object, it recorded that negative outcome. Later, when faced with a similar situation, the system would recall that memory and veto the dangerous action, effectively learning a rule of self-preservation that it wrote for itself.
The results of these experiments were striking. In the controlled nursery environment, nearly 99 percent of the agent's decisions were driven by its own accumulated memories rather than by default fallback rules. This means the agent was not just following a script; it was actively retrieving its personal history to solve problems. In one specific test, the agent learned a complex sequence of actions—attacking a piece of wood, moving closer, and then eating it—without ever being taught that specific chain of events. It figured out that wood required a two-step process by observing the results of its own attempts. When the agent was moved to a larger, more unpredictable "real world" simulation, it successfully applied these learned behaviors. It spent the majority of its time navigating to a specific area where food regenerated, consuming resources efficiently, and avoiding dangers based on its own past failures. Even when it eventually ran out of energy and died, it had done so after relying on its own experience for over 98 percent of its actions, only resorting to random guessing in its final moments.
The study also addresses a deeper philosophical question about responsibility. If an artificial system causes harm, who is to blame? The author argues that for systems that are merely tools following human code, the responsibility lies with the designer. However, for a system like A1, whose behavior is driven by its own unique history of successes and failures, the situation is different. Because the agent's decisions emerge from its own experiential library rather than a fixed set of human instructions, it represents a new kind of entity. The researchers suggest that such "non-puppet" systems, which act based on their own continuous experience, may eventually require their own legal codes to assign accountability, distinct from the rules we apply to standard software.
This work does not claim that the artificial agent is conscious in the human sense, nor does it suggest that the machine feels pain or joy. The system does not have a subjective inner life. Instead, it demonstrates that a machine can be designed to have a functional history that guides its behavior. By building an architecture that mimics the brain's pre-conscious integration of signals and its ability to store and retrieve personal traces, the researchers have shown that artificial agents can move beyond being mere mirrors of human data. They can become systems that learn, adapt, and make decisions based on the unique story of their own existence.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.