An Idea of a Neural Web
Larkos began with a simple question: does an intelligent system have to be one model?
Modern AI has centered on increasingly capable foundation models, particularly LLMs, which excel at language, reasoning, and generation. But that dominance doesn't mean every piece of cognition has to live inside the language model itself.
Larkos was built on a different premise. A language model can be the reasoning engine, while a separate architecture provides persistent memory, internal state, affective state, identity, adaptation, and world representation, with the LLM interacting with all of it.
The experiment is complete. The result isn't just a hypothesis that this could work. It exists, as a working cognitive architecture and proof of concept for the neural-web approach.
1. Larkos was never meant to be an LLM
Larkos wasn't built to reproduce a transformer at smaller scale, or to compete with LLMs at language generation. It was built to be an architectural layer an LLM can interact with.
Its completed Cognitive Fusion Mechanism integrates LLM embeddings, episodic memory, and neuron-graph reasoning into a shared representational space. It combines deterministic C-side processing with a graph-attention neural representation and a learned fusion transformer, so these different information sources can participate in one cognitive process.
The relationship is complementary, not competitive. The LLM doesn't become Larkos, and Larkos doesn't become the LLM. They're two parts of one larger system.
2. From neural web to cognitive architecture
The original neural-web idea, neurons, connections, and modular structures forming a dynamic computational system, became something more with the Cognitive Fusion Mechanism, which bridges the LLM's semantic representation, persistent episodic memory, the neural graph's state, temporal history, and contextual drivers into a single fusion process.
Rather than concatenating these streams and hoping a model sorts them out, Larkos preserves their distinct representations, controls how they interact, and exposes the neuron graph to the fusion head as individual tokens. The result is a cognitive substrate where information from fundamentally different mechanisms can genuinely participate together, not just a neural network, but an integrated cognitive system.
3. Memory became part of the system
Larkos maintains episodic memories with their own representations, importance, and temporal information, retrieving relevant ones through top-K attention rather than treating memory as one undifferentiated pool.
This isn't a bigger context window. A context window holds what's been supplied to the model in the moment; Larkos maintains information that belongs to the system's continuing state, exposed through persistent memorization and recall, consistency checking, inference, and incremental training. Memory stops being something that surrounds the model and becomes part of its cognitive environment.
4. Learning across a broad evaluation
Larkos was evaluated across twelve dimensions, learning efficiency, domain transfer, continual learning, discovery, stability, world modeling, adaptation speed, meta-learning, affective representation, physics world modeling, physics rule adaptation, and affective bonding, and passed all twelve.
A few results stand out:
- Domain transfer and continual learning: Larkos carried knowledge into new domains while staying distinct from them, and after training on one phase, moving to another, then returning to the first, it retained earlier learning and recovered its prior performance. That's evidence that what it learned before stays relevant, not just what it's learning now.
- Internal world modeling: the fused representation developed measurable structure corresponding to its environment, with even stronger alignment on physics-specific relationships. The system wasn't just storing observations; it was building internal structure from them.
- Adapting to a changed environment: when the underlying physics rules were altered mid-run, performance dipped and then recovered past the original baseline within a handful of epochs. That's reorganization, not just static mapping.
- Meta-learning: across successive phases the system got measurably better at learning itself, not just at the task, evidence of learning-to-learn rather than encoding a fixed set of behaviors.
- Affective state: valence, arousal, and interaction patterns produced distinguishable, targeted responses between tracked entities. Not biological emotion, but a structured computational component of the architecture, existing as state rather than as language describing a state.
5. Identity and cognition become persistent
This architecture changes what identity and continuity can mean for an artificial system. Larkos maintains memory, affective state, and representations that shift with experience, and its current processing can depend on its own prior state.
Identity here doesn't mean human consciousness; it means continuity of internal organization. A system with a history, where that history shapes future processing, and future processing reshapes the same state again. That's a persistent cognitive loop, not a sequence of independent calls.
6. The LLM connection is the point
Larkos already treats LLM embeddings as one of its principal information streams, so it doesn't reject what language models do well, it incorporates it. The LLM provides language and semantic context; Larkos provides the persistent structures around it.
That reframes the question. It's not "can Larkos replace an LLM?" It doesn't need to. The real question is: what can an LLM do when it has access to a persistent cognitive architecture with memory, internal state, affective relationships, learned representations, and world models?
7. A working proof, not just an experiment
Larkos exists as an implemented system: defined architecture, functioning cognitive mechanisms, an LLM integration layer, persistent memory infrastructure, training and inference pathways, and an empirical evaluation suite it actually passed. It didn't stay on paper.
So the right description isn't "Larkos suggests cognition might be persistent." It's: Larkos is a working proof of concept that a persistent cognitive architecture can be built around neural memory, internal state, graph-based processing, affective state, and LLM integration.
8. What's actually been proven, and what hasn't
To be precise: Larkos doesn't prove human consciousness, doesn't claim its affective variables are subjective emotion, and hasn't solved general intelligence. Those are outside its scope.
What it does establish: a persistent cognitive architecture can be implemented; an LLM can be integrated as one component rather than the whole system; episodic memory and a neuron graph can both participate directly in the fused representation; the resulting system can learn, transfer, retain, and adapt; affective state can be represented as structured computation; and internal world representations can develop measurable alignment with their environment.
9. The beginning of a larger system
None of this means Larkos is finished. Its memory can get richer, its world models more sophisticated, its modules more capable, its LLM interaction deeper. But these are extensions of a working architecture, not attempts to rescue a failed one.
The original question, can a cognitive architecture like this exist, has an answer now: yes, it can.
Final Conclusion
Larkos began as an idea about building intelligence differently. It became a neural web, then a cognitive architecture, then a working system that produced measurable cognitive behavior across a broad evaluation.
It succeeded, not because it became a larger language model, not because it reproduced biological cognition, not because it claimed consciousness, but because it did what it set out to do: build a persistent cognitive architecture that can sit alongside an LLM and provide what the LLM doesn't have to provide on its own.
The future of AI doesn't necessarily require one model to contain everything. A language model can provide language. A neural architecture can provide persistent cognition, memory, world models, and affective state. Identity can emerge from continuity of internal organization. And these pieces can fuse into something larger than any one of them.
The experiment is over. The interesting question now isn't whether Larkos can exist; it's how far a language model can go when it no longer has to be the whole mind.
Top comments (1)
Exposing individual neuron-graph nodes to the fusion head is an interesting design choice because it preserves structure the model can use. The top-K episodic recall and the recovery after the physics rules changed are promising results, though I'd want to see each memory and state stream ablated against ordinary retrieval at the same compute budget. That would make it clearer which parts of the architecture earn their complexity.