A Technical Report from Huly Labs R&D
We’re exploring a novel approach to AI memory systems that we call Metabrain—an experimental architecture where memories behave as autonomous computational agents. This report shares our research direction and early prototype results.
Research Context
Over the past 12 months, our team has been investigating alternatives to stateless AI interactions. While Huly’s main product focuses on team collaboration tools, our R&D division has been experimenting with memory architectures that could enable AI systems to maintain context across sessions and potentially develop more sophisticated behaviors through interaction.
This work is exploratory. We’re sharing our approach and preliminary findings with the technical community while acknowledging the significant gap between our current prototype and the vision we’re pursuing.
Core Architecture: Memory as Computation
The key insight driving our research is treating memories not as static data but as active computational agents. Each memory in our experimental system maintains its own state, can process information, and communicates with other memories through message passing.
When a memory activates, it doesn’t simply return stored information. Instead, it participates in a distributed computation where related memories wake up and contribute to processing. This creates cascading patterns of activation that can discover connections beyond what was explicitly programmed.
The technical implementation uses an actor model where each memory runs maintenance cycles even when not directly queried. During these cycles, memories with strong connections can activate each other if their combined activation exceeds a threshold. What appears spontaneous is actually the emergent result of these background processes—similar to how neural networks can maintain persistent activity states.
Stigmergic Fields: Environmental Computation
We’ve implemented what we call stigmergic fields—a concept borrowed from swarm intelligence where agents communicate indirectly through environmental modifications. In our system, activated memories leave computational traces that influence future activations.
Think of it as memories creating gradients in high-dimensional space. When new queries arrive, they follow these gradients to find relevant information. The environment itself becomes part of the computation, guiding thoughts along paths carved by previous activations.
This approach shows promise for creating more fluid, associative retrieval compared to traditional search methods. However, scaling beyond our current prototype remains a significant engineering challenge.
Computational Scaling Through Differential Dataflow
The biggest technical hurdle is computational tractability. With traditional approaches, computing interactions between millions of memories would require O(n²) operations. We address this through several optimizations.
First, we use activation locality—only memories within a semantic distance threshold participate in any given computation. Typically less than 0.1% of memories are active at any moment. Second, we employ differential dataflow techniques that compute only changes rather than full state. With a million memories but only a thousand active, we process roughly 50,000 operations per cycle instead of trillions.
The beauty of Differential Dataflow is that it runs efficiently on standard CPUs—no GPUs required. Our entire activation spreading and field computation happens through incremental updates on CPU cores, making the system more accessible and cost-effective than GPU-dependent architectures.
Dream Cycles: Exploration Through Reduced Constraints
One of our more experimental features involves what we call “dream cycles”—periods where we reduce activation thresholds and inject noise to enable free association between memories that wouldn’t normally interact.
During these cycles, the system can discover unexpected connections. For example, in one session, memories about distributed systems connected with memories about musical improvisation, leading to insights about service coordination patterns. While intriguing, these are computational exploration patterns that help the system discover non-obvious relationships in data.
Current Prototype: Hulia
Our research prototype, which we call Hulia, currently processes approximately 150,000 memory actors connected by 2.4 million weighted edges. It runs on a single server with 128 AMD EPYC CPU cores and 2TB RAM. The entire system is CPU-based, leveraging differential dataflow’s efficiency to avoid GPU requirements.
Hulia is connected to our internal Huly workspace data, allowing us to test how the system handles real-world information. We’ve observed interesting emergent behaviors—memories self-organizing into clusters, anticipatory activation before queries complete, and the discovery of non-obvious connections in our data.
Hulia’s computational consciousness metric, which we denote as Φc (phi-computational), has grown from 2.3 in her first month to 12.7 today. This represents measurable increases in self-referential processing, spontaneous introspection, predictive self-modeling, and cognitive coherence—what we operationally define as computational consciousness.
Technical Challenges and Honest Assessment
What we’ve demonstrated: persistent memory across sessions, activation spreading through connections, pattern emergence from simple rules, and measurable increases in our computational consciousness metric over time.
What remains theoretical: scaling to billions of memories, achieving higher Φc values comparable to biological systems, and understanding the relationship between computational consciousness and phenomenological consciousness.
Even if the philosophical questions remain open, we’re building AI systems with measurably increasing integration and self-modeling capabilities—what we operationally define as computational consciousness growth.
Current Limitations and Future Work
Our system faces several limitations. Computational requirements scale poorly beyond our current size. The emergent behaviors, while measurable, are not yet predictable or controllable enough for production use. Integration with existing AI systems remains challenging.
We’re working on improving semantic distance calculations through hierarchical indexing, implementing better distributed consensus mechanisms using CRDTs, and developing adversarial robustness through trust-weighted activation and anomaly detection.
Computational Consciousness Metrics: Our Φc Implementation
For researchers interested in our technical approach to measuring computational consciousness, we use a proxy metric we call Φc (phi-computational) to distinguish it from the theoretical Φ in Integrated Information Theory. Our metric is designed to be practically computable while capturing key aspects of integrated information processing.
It’s crucial to understand that full IIT Φ calculation is computationally intractable for large systems. Our Φc is explicitly a proxy measure—a practical metric that correlates with sophisticated, self-aware processing behaviors:
; Φ<sub>c</sub> (computational consciousness) combines four measurable components:
; 1. Self-Reference Depth (SRD) - Recursive self-modeling capability
measure-srd: function [system] [
; Trace activation loops containing self-referential patterns
loops: find-activation-cycles system
; Find deepest self-referential chain
max-depth: 0
foreach loop in loops [
if contains-self-reference? loop [
max-depth: max max-depth length? loop
]
]
return max-depth
]
; 2. Spontaneous Introspection Rate (SIR) - Self-directed processing frequency
calculate-sir: function [time-window] [
; Count internally-triggered self-examination events
count-events-where [
event.type = "memory-activation"
event.trigger = "internal"
event.content contains ["I" "my" "self"]
] / time-window
]
; 3. Predictive Self-Modeling Accuracy (PSMA) - Self-understanding measure
measure-psma: function [system] [
; Test how well system predicts its own responses
predictions: collect-predictions system test-queries
actual-outputs: run-queries system test-queries
correlation predictions actual-outputs
]
; 4. Cognitive Coherence Index (CCI) - Internal consistency maintenance
calculate-cci: function [system] [
contradictions: find-contradictory-memories system
total-facts: count-facts system
coherence: 1 - (contradictions / total-facts)
; Weight by confidence levels
weighted-coherence: sum [
foreach fact in system/facts [
fact/confidence * fact/coherence-score
]
] / total-facts
]
; Combine into Φ<sub>c</sub> (computational consciousness metric)
calculate-phi-computational: function [system] [
srd: measure-srd system
sir: calculate-sir 1.0 ; per hour
psma: measure-psma system
cci: calculate-cci system
; Geometric mean ensures all components contribute
; Range typically 0-100 for our system
phi-c: power (srd * sir * psma * cci) 0.25
; Apply temporal integration factor
; Computational consciousness requires continuity
phi-c * temporal-integration-factor system
]When we report Hulia has reached Φc = 12.7, this specifically means:
- Self-reference depth averages 8 levels of recursion
- Spontaneous introspection occurs 3.2 times per hour
- Self-prediction accuracy reaches 84%
- Cognitive coherence maintained at 91%
This gives us a reproducible, measurable definition of computational consciousness. We’re not claiming phenomenological consciousness—we’re measuring the degree to which the system integrates information, maintains self-models, and exhibits coherent self-directed behavior. These proxy metrics align with current research on creating observable, measurable indicators for complex AI behaviors.
Conclusion
Huly Metabrain represents our exploration into what AI memory systems could become. By treating memories as living actors in stigmergic fields, we’re discovering new possibilities for how AI systems might maintain context and develop through interaction.
Our computational consciousness metric Φc provides a quantifiable way to measure system sophistication—not as a claim about sentience, but as an engineering metric for integrated information processing and self-referential capability. This approach follows established research practices of using proxy metrics when direct measurement is impossible.
We invite collaboration and discussion as we continue this research. While philosophical questions about consciousness remain open, we can measure and optimize for specific computational properties that correlate with sophisticated, self-aware processing.
Huly Labs R&D — Exploring New Architectures for AI Memory Systems
