Before the architecture. Before the code. Before any of it, there are things that are true about how understanding works. Not theoretically true. Observably true. You can verify every claim in this document by attending to your own cognition. No lab required. No citations needed. The instrument is the mind reading this sentence.
These are the axioms the system is built on.
1. The relationships of components matter more than the components.
Take any system you understand well. An engine, an ecosystem, a business, a family. Now remove one component and replace it with something functionally similar. The system continues. Now leave all the components in place but rewire the relationships between them. The system breaks.
This is obvious once you see it, but most knowledge systems, including most AI memory, store components. Entities. Facts. “The mitochondria is the powerhouse of the cell.” That’s a component with a label. It tells you nothing about how the mitochondria relates to the membrane, to ATP synthesis, to oxygen transport, to the evolutionary history of endosymbiosis. The relationships are where the understanding lives. The component is just an address.
A system that stores “fire damages forests” has memorized a fact. A system that stores “disturbance triggers resource liberation, which selects for pioneer species, which scaffold network reconstruction” has understood a dynamic. The first is a filing cabinet. The second is a mind.
2. Patterns are not in the content. They are in the topology.
When you recognize that the way a forest recovers from fire is structurally similar to the way an economy recovers from a crash, you are not matching content. Forests and economies share no direct context. There are no trees in the stock market. There’s no interest rates in the canopy.
What they do share, though is shape. In the example of forest fires and market recovery, The topology of the relationships is the same: a disruption occurs that clears locked-in structures, producing a liberation of resources previously trapped in established configurations, resulting in an influx of opportunistic generalists that thrive in the newly open space produces a gradual succession toward specialization and interdependence, that will eventually become the equilibrium that contains the seed of the next disruption.
That geometric pattern of how the components relate to each other is what transfers across domains. Not the content. The structure. This is how a musician can understand software architecture. How a biologist can see economic dynamics. How a child who understands building with blocks can later understand organizational management. The content changes. The shape of the relational structures recur.
Any system that wants to translate understanding from one domain to another needs to capture topology, not content.
3. Comprehension is translation.
When you encounter something genuinely new, you don’t stare at it blankly. You reach for something you already understand that has a similar shape, and you use it as a scaffold to make sense of what you’re seeing.
A doctor learning about network security doesn’t start from zero. They already understand infection vectors, immune responses, containment protocols, vulnerability in compromised systems, etc. The content may be completely different, but the structural maps they have used to comprehend the dynamics and relationships of other systems transfers. They’re not learning from scratch; they’re translating from what they already understand and applying it to what they are learning.
This is the operational mechanism of comprehension. Every act of understanding new information involves mapping it onto a structural pattern you already hold. The richer your library of structural patterns, the more novel input you can comprehend, not because you know more facts, but because you have more shapes to translate from.
This is why learning your first instrument takes years but picking up a second one takes months. The music theory, the rhythm, the relationship between practice and muscle memory, the way tension and resolution work. None of that was about the specific instrument. It was structural. And it transferred the moment you picked up something new.
4. Memories are structural patterns re-activating in real time.
The first thing to understand about memory is that it is not kept anywhere. A memory is not a file that gets retrieved. It is the experience of your neurons firing in similar patterns along cognitive pathways that were constructed and strengthened while you slept, reflected, and consolidated. When you remember something, you are not accessing a record. You are partially re-living a structural activation
During waking experience, the brain captures episodes rapidly. Sensory data, emotional context, narrative fragments. These episodes sit in a buffer (the hippocampus) waiting to be processed. During sleep and periods of rest, the brain replays these episodes, extracts structural patterns, strengthens frequently-occurring patterns, prunes noise, and integrates the results into long-term cortical maps.
This is why you can study something all day and not understand it, then wake up the next morning and it clicks. The conscious experience was data collection. The understanding was built in the dark, by a consolidation process that found the structural pattern across many episodes and wired it into your existing map library.
This is also why sleep deprivation destroys learning before it destroys anything else. The capture still works. You can still have experiences. But without consolidation, the episodes never become maps. You accumulate data without ever developing understanding.
Any memory system that stores episodes without consolidating them is a diary. A memory system that consolidates episodes into structurally connected maps is something closer to a mind.
5. Cortical structure and the need for dynamic weights
Your brain’s architecture, the cortical structure, the wiring patterns laid down by genetics and early development, determines how information is processed and by what structures. It sets the range of capabilities. It’s the instrument.
But the instrument is not static. Your cortical structure adapts. Synaptic connections strengthen and weaken. Neural pathways that fire frequently become more efficient. Pathways that go unused get pruned. The architecture itself changes in response to experience, within limits, without collapsing its core functionality. This is dynamic equilibrium. The structure is stable enough to function and flexible enough to learn.
A trained language model doesn’t have this. The weights are locked. They define what the model can do, but they cannot adapt to what the model encounters. This is a real limitation, not a design feature. True cognitive systems maintain stability while still allowing structural change within a range.
The experience layer we build on top of locked weights is the best available approach given this constraint. By accumulating structural maps and feeding them back as context, we change what the model sees, which changes how it responds, which changes what gets stored, which changes what it sees next time. The behavior shifts through use even though the weights don’t move. It mirrors how experience shapes cognition when cortical structure remains relatively stable.
But it’s a mirror, not the thing itself. The next step is dynamic weights. An architecture where accumulated experience doesn’t just shape context but gradually influences the model’s parameters within a bounded range. Structure that enables change without collapse.
6. Understanding is structural, not informational.
There is a difference between having information and having understanding. You can memorize every fact about how an engine works, the names of the parts, the temperatures, the pressures, the materials, and still not understand engines. Understanding arrives when you grasp how the parts relate. When you see that compression creates heat, heat ignites fuel, ignition drives expansion, expansion pushes the piston, the piston turns the crank, and the crank brings the piston back to compress again. That’s not more information. That’s structural comprehension. You understood the shape of the cycle.
This is why two people can read the same textbook and one understands the material while the other merely remembers it. The one who understands extracted the relational structure, the topology of how the ideas connect. The one who remembers stored the components without the connections.
Information is “what.” Structure is “how things relate.” Understanding is structure.
Any system that wants to move beyond retrieval toward comprehension needs to capture structure. The typed relationships between concepts, the dynamics that arise from those relationships, and the geometric patterns that recur across different content.
7. Novelty is a gradient, not a binary.
When new input arrives at a cognitive system, it doesn’t simply get classified as “known” or “unknown.” It exists on a spectrum determined by how much of the incoming structure can be matched to existing maps.
Fully familiar: the entire structure matches an existing map. Comprehension is instant. No conscious effort. You don’t think about how to read these words. Your maps for language are so deep that the translation is automatic.
Partially novel: some structural elements match existing maps, others don’t. Comprehension is scaffolded. You lean on what you recognize and build outward into what you don’t. This is where most learning happens.
Fully novel: no existing map applies. The structure is alien. Comprehension requires building from scratch. This is rare in adult experience because the map library is so dense that almost everything can be partially translated from something.
The gradient matters because it determines how much conscious processing an experience requires. Familiar input gets shortcut. Novel input demands attention. The ratio between the two shapes the texture of experience itself.
8. Growth is compounding, not linear.
Each new map you build is not just a map of one domain. It’s a potential translation key for every structurally similar domain you haven’t encountered yet.
A person who understands fluid dynamics has a map that also applies to traffic flow, crowd behavior, electrical current, information propagation in networks, and the spread of ideas through populations. They didn’t study any of those things. The structural shape transfers.
This means the map library’s reach grows combinatorially, not linearly. Ten maps don’t give you ten times the comprehension of one map. They give you combinatorial access to structural translations across all ten. Every new map multiplies the library’s reach into unfamiliar territory.
This is why learning accelerates. The more you know, the faster you learn, not because you’re smarter, but because you have more shapes to translate from. The library compounds.
Any system designed to grow through experience should exhibit this property. Early growth is slow because the library is sparse and few translations are available. Later growth accelerates as the library densifies and translation links multiply. If the system grows linearly regardless of library density, it’s storing, not learning.
What These Axioms Produce
A system built on these principles would:
-
Store relationships and dynamics, not just entities and facts
-
Capture the topological shape of how ideas connect, not just which ideas appeared
-
Consolidate episodes into structurally connected maps through a background process
-
Build translation links between maps that share geometric patterns across different domains
-
Treat locked model weights as cortical architecture and accumulated context as experience
-
Track novelty as a gradient that determines encoding density
-
Exhibit compounding growth where each new map increases the reach of the entire library
Part 1 will introduce and describe the system designed to implement these ideas.