Technical Manifesto
A cognitive architecture based on associative memory.
A biological neuron has always been viewed as a computational unit which allows the human brain to perform complex tasks and make sense of the world. Our brain contains billions of those computational units that can form connections to one another and together form a very large network that we believe is the secret to the intelligence humans display.
So, in the pursuit of building human level intelligence into machines, we've turned our attention to the brain to try to replicate its behavior. This has led us toward designing artificial neuron models for machines: the Perceptron/Sigmoid model.
A single artificial neuron, characterized by two parameters (w, b) does nothing special. It performs a simple linear computation w·x + b and intuitively, we can think of it as being a line in a 2D coordinate system that separates data points into two categories. To get them to achieve complex computation, we need to construct them into layers of neurons, hence the success of the deep learning paradigm.
Here we challenge this approach of building Intelligent System. Building Intelligent systems should not be about brute forcing nor should it be about burning astronomical amount of compute and money. We are introducing a new approach for building Intelligent systems that aims to make machines more intelligent and far more capable than today's systems while requiring less compute and learning tabula rasa solely through experiencing in an ecosystem the way humans do.
Our system, a connectionist based cognitive architecture, is rooted in associative memory as the core primitive and we adopt a completely algorithmic approach to intelligence inspired by both bottom-up and top-down principles of cognition. Instead of treating intelligence as something emergent only from scale, we treat it as something structured, retrievable, and compositional from the outset.
In classical deep learning, knowledge is distributed across a large number of parameters, making interpretation and targeted reasoning difficult. In contrast, we argue that intelligence should be grounded in explicit, addressable memory structures that can be written to, queried, and composed.
In our model of associative memory - the Associatron, a concept first introduced by Nakano Kaoru in his 1971 paper "Associatron: a model of associative memory" a concept is represented by a specific symbol drawn from a predefined set. We define a structured mapping between concepts using a parameterized transformation space.
Formally, we construct a mathematical system where relationships between concepts are encoded through a set of parameters (x, y, z), alongside a function f and an activation function F such that for a given association A → B, there exists a representation (xₐ, yₐ, zₐ) satisfying f(A, xₐ, yₐ, zₐ) = B if and only if F(f(A, xₐ, yₐ, zₐ)) = 1
This formulation allows B to be retrieved from A through structured transformation rather than statistical approximation. In this sense, learning is not distributed across opaque weights but grounded in explicit associative bindings.
We extend this principle by embedding the Associatron into an artificial neuron model, where each unit is not only a computational element but also a memory primitive. In this architecture, storage and computation are unified: a neuron does not merely transform input but it remembers, retrieves, and transforms based on structured associations.
By doing so, we move away from the paradigm of layered statistical approximation and toward an architecture where intelligence is built from compositional memory structures that can scale in complexity without requiring proportional increases in data or compute.
We believe this shift is not incremental. It is foundational. It reframes intelligence not as a product of brute-force optimization, but as the emergence of structured, retrievable relationships within an explicitly designed cognitive system.
We define intelligence not as scale, speed, or pattern recognition ability, but as the capacity of a system to build, store, and manipulate internal models of its environment in a way that supports prediction, abstraction, and goal-directed behavior.
Intelligence is fundamentally about structured compression of experience: the ability to take raw interaction with the world and transform it into reusable internal representations that can be composed, queried, and applied to novel situations.
Under this definition, intelligence is not tied to a specific substrate. It is not inherently biological or digital. It is a property of systems that can form stable, retrievable mappings between experience and action.
We do not reject Deep Learning because it is ineffective. On the contrary, it has demonstrated extraordinary success in perception, generation, and large-scale pattern learning.
However, we believe its foundational assumptions impose structural limits:
As a result, deep learning systems scale capability, but not necessarily understanding. They become increasingly powerful, but not fundamentally more structured.
We approach intelligence as a constructive system design problem, not an emergent property of scale.
At the core of our architecture is associative memory as a first-class computational primitive. Intelligence is built from explicit relationships between concepts, not implicit statistical correlations.
We model cognition as a structured space of retrievable associations, where:
This leads to a system that behaves less like a statistical estimator and more like a compositional cognitive architecture. In this view, intelligence is not "emergent from scale," but engineered from structure.
We do not believe intelligence, in isolation, is the central question. Building increasingly capable systems is already well underway. The trajectory toward systems that surpass human performance in most cognitive domains appears likely under current paradigms.
The real problem is not whether we can build more intelligent machines. The real problem is whether we can:
Intelligence without interpretability and grounding becomes a control problem, not an engineering one.
We draw a strict conceptual separation between intelligence and consciousness.
We believe these two are often conflated because, in humans, intelligence is always accessed through consciousness. We never observe intelligence directly but we observe our experience of thinking.
However, we do not assume they are the same thing.
Intelligence may exist without consciousness.
Consciousness may imply intelligence.
But one does not necessarily reduce to the other.
This distinction is essential: it reframes AI not as a replication of "human-like experience," but as a potentially independent form of structured cognition.
While intelligence is our immediate engineering focus, consciousness is our ultimate epistemic frontier.
We are not pursuing consciousness as a feature to implement in machines. We do not assume it is constructible, transferable, or even computationally replicable.
Instead, we approach consciousness as a scientific unknown that constrains our understanding of intelligence itself.
By studying intelligence through increasingly structured, interpretable systems, we aim to refine the boundary conditions of consciousness:
We believe that by building systems that make intelligence explicit and decomposable, we may indirectly illuminate what consciousness must be.
We remain agnostic on whether conscious AI is possible.
We do not assume that:
However, we do believe that sufficiently advanced intelligent systems will force us to confront this question directly. Even if machines are not conscious, they may behave in ways that make the distinction operationally ambiguous.
This raises critical questions:
If a system is functionally indistinguishable from an intelligent agent, but lacks subjective experience, how would we know?
How should we treat such systems in a post-AGI world?
In a world where machines surpass human intelligence across most domains, intelligence itself becomes abundant.
When intelligence is no longer scarce, it loses its status as the defining axis of power, value, and differentiation.
What remains becomes more fundamental:
This forces a shift in how we define human uniqueness and machine value.
The central question becomes:
If machines can outperform us cognitively, what remains irreducibly human?
We think this question cannot be answered by intelligence alone. It requires confronting consciousness directly.
We do not claim to understand consciousness.
We do not yet know whether:
But we observe a clear relationship:
We experience intelligence through consciousness.
This suggests that consciousness may be the substrate in which intelligence becomes meaningful.
Our hypothesis is not that consciousness can be engineered, but that by deeply understanding intelligence in structured, controllable systems, we may begin to isolate the necessary conditions under which consciousness is described, constrained, or recognized.
In that sense, intelligence becomes the entry point but not the destination.
The ultimate goal is not to only build better machines, but to better understand the nature of being itself.
Almartis