Proxima - Facts, Goals, and the Boundary of Life

Touch your arm, and you feel it right there. Yet the sensation does not arise in your arm. It arises in your head. This system can be fooled: under certain conditions, merely seeing a touch can trigger the sensation of being touched even though nothing has made contact with the skin. There are many examples like this, and they all reveal the same thing: what we experience as reality does not simply happen outside us. It is constructed in our heads. Our nerves supply the signal.

None of this is new. Yet the more I work with LLMs and AI, the less this thought remains philosophy and the more it grows into a persistent question. What separates a conscious decision from an unconscious one? How can an inner model of the world be represented? And what, exactly, separates a chatbot that feels alive from a living being? People have long asked how conscious life should be defined. Today, however, an entirely new perspective on that question has opened up. Proxima grew out of my attempt to answer these questions as a system.

Let us begin outside. An event in the world has happened. Yet it never reaches us directly, only as an observation. An observation of this kind is a Fact (F). It records what a source observed at a particular time. If the observation later proves incomplete or wrong, it is not rewritten after the event. A new fact is added. Without the nerve in your arm, the touch would never arrive in your reality at all. A fact is therefore nothing universal. It is always bound to an observer.

Many facts reach us through our senses: we feel, hear, see and taste. But what we make of them depends heavily on our conditioning. This includes our experiences, our present state and our brain's interpretation. A fact is the pure, uninterpreted observation. We are never neutral, nor can we be, because every observation is processed through learned patterns. “Learned” here means two things: acquired genetically over long stretches of time, and taken on actively through experience over the course of our lives. The processing is shaped. The record is not.

What do all living beings have in common? Not chemistry alone. That would confine life to the narrow circle of our own organic compounds. My proposal is an operational one: life manifests wherever a system autonomously maintains or pursues states that support its continued existence. This directedness is the second foundational concept. The Goal (G). One could also argue that what happens here is a local, directed reversal of entropy. What matters is not the reversal itself — hurricanes and crystals manage that too — but that a system willingly expends energy on it. Life orders, and it orders in a direction.

From bacteria and plants to complex animals, living beings develop strategies for reaching favorable states. Their degree of freedom and the complexity of their goals vary enormously, yet the principle is the same. A few basic directions are usually given, such as self-preservation and reproduction, and these yield subgoals such as obtaining energy. The virus remains an interesting boundary case: outside a host cell, it cannot pursue such states autonomously. The usual objection — that it draws its energy from outside — does not carry far, however; every living thing does that. What matters is something else: a virus pursues a state of its own, it merely borrows someone else's machinery to do so. If you allow that, it is alive — and the boundary runs not at metabolism but at having a goal of one's own.

Here we encounter a first difference from an ordinary program. Without an inner goal of its own, a program has only the task supplied by a user or another system. Its goal is not intrinsic. It depends entirely on external input.

Where a goal exists, it becomes a measure: observations can be evaluated by whether they move a system closer to or farther from a desired state. In the simplest forms of life, this is chemical regulation without any experience at all. A starving bacterium does not “feel” anything. It reacts. In more complex nervous systems, such evaluation can become reward or aversion, pleasure or stress. This does not reduce every feeling to reproduction. It means only that evaluation gives behavior a direction. Without a concrete goal there is no measure to define, and without a measure there is no evaluating your own actions.

How do we bridge facts and goals? In simple forms of life, the two are directly connected. A bacterium such as E. coli essentially has two modes of movement: swimming straight ahead or randomly changing direction, known as “run and tumble”. What matters is not the absolute concentration of nutrients but how it changes over time. When the sensed concentration rises, the bacterium extends its straight runs. When conditions develop unfavorably, the probability of a change in direction increases. Fact in, measured against earlier facts and a favorable state, action out. No conscious decision is required. Regulation is enough.

But we can do more. We can learn. We can abstract. This is the third concept: Abstraction (A). An abstraction is formed from facts by extracting their pattern and what is essential about them. When a child touches a hot stove, what exists at first is an observation: surface touched, pain felt. Only when the child derives the general relationship that hot surfaces can cause burns from this and similar observations does an abstraction arise.

Abstractions therefore lie between facts and goals, and they are no longer found in every living being. Both foundations are necessary. Without goals, there is no reason to abstract. Without facts, there is no basis. This also helps explain why LLMs often become dramatically more useful when connected to tools. Tools expand what an agent can observe and affect. Yet more data alone does not make it more intelligent. Value emerges only when it can classify observations in light of its goals and preserve the insights it gains.

Yet we are not made up of facts, goals and abstractions alone. The most important piece is still missing: Perspective (P). Think back to the touch on your arm. Mechanically it is the same event. And yet it makes an enormous difference who touches you, and where. The difference lies not in mechanics alone, but also in your perspective, which places the person and the context.

Perspective influences how we relate goals and facts and how we form new abstractions. Conversely, our abstractions shape our perspectives and form a self-reinforcing loop. I deliberately say perspective rather than personality: we carry different perspectives on different subjects. In this model, personality can be understood as their weighted combination.

This loop offers a model for two familiar phenomena. First, why breaking bad habits is so difficult: a learned perspective favors precisely the abstractions that confirm it. Second, how opinions can become radicalized. Social media is built to hold our attention, and it does so particularly well when it aligns with existing perspectives. More confirming information, more uniform abstractions and sharper perspectives form a closed feedback loop.

The Machine Without Memory

These four concepts describe a surprisingly large part of our behavior. The subconscious resembles a statistical machine. Decisions are made on the basis of our conditioning, whether biologically given or acquired through experience. Decisions create new facts. Facts become perspective-shaped abstractions. These are measured against goals. Then the loop begins again.

And now the provocative question: what happens to our experience of continuity when hardly any new memories remain, when impressions merely pass through a statistical machine? Part of the answer can be seen in rare, well-documented cases of severe damage to the medial temporal lobe. The famous patient H.M. underwent bilateral tissue removal in 1953, including parts of the hippocampus. Afterwards, he could form hardly any new declarative long-term memories, while older memories and certain forms of learning remained partly intact. What happens to our inner model of the world when new experiences can no longer become lasting abstractions and perspectives?

I do not ask this by accident. An isolated LLM call is stateless: the model processes the supplied context through a pretrained network, but does not itself form a new durable memory between calls. The parallel is not poetic but structural. H.M. could not say what he had done an hour earlier, yet he grew better at tasks he had no memory of practicing. Seen from the outside, that is uncomfortably close to what an agent with tools and without memory achieves. In both cases the same organ is missing: the path from experience into lasting memory. Applications can present earlier messages again, yet no memory of its own grows inside the model. The loop is missing. No facts that remain. No abstractions that grow. No perspective that sharpens. What happens to such a machine when you give it exactly that?

The final foundation for agentic systems

Proxima as a System

Proxima is precisely this loop as software: a typed, durable memory substrate for agentic systems, built on Facts, Abstractions, Perspectives and Goals. It gives the stateless model a structured memory. And it is more than that — it is a philosophy of how independently acting systems come about: from memory through to the question of when an agent should act on its own.

It is governed by a few fundamental rules:

These rules are not mere recommendations. They form Proxima's implemented core, and with it the basis for the sensible, purposeful use of agents. Of course the other half is needed too, the intelligence, and this is exactly where LLMs come in. The model represents the subconscious; Proxima serves as the basis for active learning. Only through Proxima does the LLM become an instance that can keep learning over the long run. And this holds regardless of how the LLM itself develops: the knowledge, everything learned, remains, no matter which model is used. The agent is “woken” for a purpose and can then act, with full access to the relevant memories.

You will find every further technical detail in the open-source repository (Apache-2.0): github.com/Aquilo-Solution-S/Proxima. I welcome your feedback, especially your perspectives.