Status

Proposed

Context

Alexei Kabanov describes an AGI model grounded in continuous attention movement. The core thesis: the atom of cognition is not an object, but a trajectory — a transition from one distinction to another.

The model specifies a 4-layer attention hierarchy, a gierarchical competition structure, a system broker, peripheral drivers, and a self-consciousness loop. It encodes 14,000+ generalizations, consolidated during dream state, with a binary tree of 400 core concepts.

The Integral stack (AME, ITC, agent memory) requires a cognitive substrate. Question: Can Kabanov's model serve this role?

Decision

Integrate Kabanov's continuous attention model as the cognitive substrate for AME and ITC.

• AMEs use the 4-layer hierarchy for perception routing.
• ITC uses the system broker and gierarchical competition for autonomous reasoning.
• Agent memory uses peripheral drivers for transparent, reverse-traceable state encoding.

Options Considered

Positive

• AME gains formal cognitive grounding in attention movement theory.
• ITC gains gierarchical competition structure for autonomous reasoning.
• Agent memory gains transparent, traceable encoding via peripheral drivers.
• Community stakeholders gain auditability — every decision traces to sensor input.
• Dream state consolidation maps to memory hygiene during idle cycles.
• Pochemuchka curiosity drives map to active hypothesis testing.

Negative

• Kabanov model is not yet formalized in code — requires implementation.
• Peripheral driver transparency adds encoding overhead.
• Gierarchical competition may create latency in time-sensitive IoT scenarios.
• Dream state consolidation requires idle-cycle management.

Risks

• Overfitting to single cognitive theory — keep AME interface abstract.
• Generalizations may not scale to diverse sensor modalities (DHT22 vs soil moisture vs solar).
• System broker becomes single point of failure if not replicated.

Core Insight: Cognitive Atom Is a Transition

Traditional AI treats perception as object recognition. Kabanov treats it as trajectory navigation — attention moves through phase space along learned routes.

This aligns with Active Inference: the brain is a prediction machine, not a classifier. AME (Active Model Engine) already uses this framing. Kabanov provides the formal trajectory model AME needs.

4-Layer Attention Hierarchy

Layer 1 — Peripheral: raw sensorimotor tokens. Layer 2 — Semantic: character-level feature binding. Layer 3 — Contextual: features plus learned context. Layer 4 — Transcendent: values and purpose drive routing.

This maps cleanly to AME processing stages:

• Peripheral = sensor inputs (DHT22, soil moisture, current State)
• Semantic = feature extraction (Bayes Net node scoring)
• Contextual = causal inference with prior context
• Transcendent = goal-state evaluation for action selection

Gierarchical Competition

Multiple super-generalizations compete for attention resources. The system broker ranks requests and builds optimal navigation routes.

This directly maps to ITC (Independent Thinking Controller):

• Each super-generalization is an ITC reasoning branch
• The system broker is the ITC meta-controller
• Confirmation/refutation/timeout map to ITC verdict cycles

Peripheral Drivers Must Be Transparent

Every peripheral driver must encode sensorimotor state transitions. It must be reverse-traceable — every decision can be followed back to a sensor event.

This is critical for community trust:

• RegenTribe members must be able to audit why a decision was made.
• The "dark factory" (Vitali's end-to-end software factory) requires auditability.
• Transparency maps to safety-case evidence in safety-critical deployments.

Dream State Consolidation

14,000+ generalizations consolidated into 400 core concepts via binary tree. This maps to memory consolidation in agent sleep cycles:

• Periodic consolidation of AME experience into generalization tree.
• Drift detection triggers re-consolidation.
• New generalizations enter at leaf nodes and propagate upward.

Pochemuchka — Curiosity-Driven Learning

The "why-asker" generates questions in character voices. It drives active learning by probing generalizations for weaknesses.

This maps to AME curiosity drives:

• AME generates hypotheses to test against environment.
• Failed hypotheses trigger generalization restructuring.
• Success confirms trajectory stability.

Learning Bootstrap via Recursive Stories

Large texts failed to trigger stable trajectory formation. Recursive iterative stories (e.g. "The House That Jack Built") succeeded.

This is a key finding for AME bootstrapping:

• Start with simple recursive structures, not corpus ingestion.
• Build trajectory stability before introducing complex texts.
• Map to story-level learning in community onboarding flows.

Three Outcomes: Confirmation, Refutation, Timeout

Every attention move has three possible outcomes:

• Confirmation: trajectory holds, continue.
• Refutation/branch: trajectory fails, spawn new branch.
• Uncertainty/timeout: insufficient data, await more evidence.

AME already uses this pattern via Bayes Net scoring. ITC uses it for verdict cycling.

Prenotion: Trajectory History

The system broker stores history of each generalization's trajectory. This is the agent's episodic memory.

AMEs must store:

• Sequence of attention moves per generalization.
• Confirmation frequency per trajectory.
• Drift events and consolidation timestamps.

This provides explainability and auditability for RegenTribe stakeholders.

References

• Alexei Kabanov, AGI model (Russian transcript)
• Active Inference Institute — Free Energy Principle
• AME (Active Model Engine) — Integral stack design
• ITC (Independent Thinking Controller) — Integral stack design
• Integral Collective dark factory concept (Vitali)