ARIA-CM: Unified Cognitive Memory Architecture

ARIA-CM: A Unified Cognitive Memory Architecture for Artificial General Intelligence

Technical Whitepaper v1.0

Abstract

We present ARIA-CM (Adaptive Recursive Intelligence Architecture - Cognitive Memory), a novel memory architecture that unifies unconscious processing, hierarchical memory consolidation, and metacognitive self-awareness into a single coherent system. ARIA-CM introduces several groundbreaking concepts: (1) Dark Matter Processing implementing predictive coding where 90% of inputs are processed unconsciously with only prediction errors reaching awareness, (2) Seven-Layer Memory Hierarchy from working memory to soul-level identity with Bayesian routing between layers, (3) Quantum-Inspired Retrieval using superposition states that collapse only upon observation enabling zero-waste memory access, (4) Living Semantic Centroids that migrate, grow uncertainty, share knowledge, and die based on usage patterns, (5) Six-Level Hierarchical Bayesian Learning including meta-learning that learns the learning rate itself, and (6) Global Workspace Consciousness implementing Baars' theory with competition, broadcast, and ignition dynamics. Our architecture achieves human-like memory characteristics: emotional memories are stronger, surprising events are remembered, routine fades, and the system knows what it knows. We demonstrate that ARIA-CM provides the first complete cognitive memory architecture for AI systems.

Keywords: Cognitive Architecture, Predictive Coding, Bayesian Memory, Global Workspace Theory, Hierarchical Memory, Metacognition, Quantum-Inspired Computing, Artificial Consciousness

1. Introduction

1.1 The Problem

Current AI memory systems suffer from fundamental limitations that prevent human-like cognition:

• Flat Memory: All information stored at same importance level
• No Forgetting: Systems accumulate everything without consolidation
• No Emotional Weighting: Routine and trauma stored identically
• No Surprise Detection: Expected and unexpected treated equally
• No Self-Knowledge: Systems don't know what they know or don't know
• Context Window Death: LLMs lose all memory at token limits
• No Learning Consolidation: No sleep-like memory processing

These limitations result in AI systems that are powerful but fundamentally lack the memory characteristics that make human cognition adaptive, efficient, and contextually aware.

1.2 Our Contribution

ARIA-CM introduces a unified cognitive memory architecture inspired by neuroscientific principles:

1. Predictive Coding Gate: Like the brain, 90% of processing stays unconscious. Only prediction errors (surprises) reach awareness, dramatically reducing cognitive load.
2. Seven Memory Layers: From millisecond working memory (L0) to permanent soul-level identity (L6), with automatic consolidation and decay matching human memory characteristics.
3. Bayesian Intelligence: Every memory operation produces not just results but confidence intervals. The system knows when it's guessing versus certain.
4. Quantum Superposition: Memory candidates exist in probabilistic superposition until observed. Only upon query does the wave function "collapse" to load actual content - zero unnecessary database operations.
5. Global Workspace: Implements Baars' Global Workspace Theory where conscious content emerges through competition and broadcast, with ignition threshold for true awareness.
6. Metacognitive Layer: The system maintains a self-model tracking capabilities, limitations, confusion states, and learning trajectories - it thinks about its own thinking.

1.3 Design Philosophy

ARIA-CM is built on principles derived from cognitive neuroscience:

1.4 Cognitive Science Foundations

ARIA-CM synthesizes multiple established theories:

2. Architecture Overview

2.1 System Structure

ARIA-CM consists of four primary subsystems integrated through a unified consciousness bridge:

                    ┌─────────────────────────────────────┐
                    │         UNIFIED CONSCIOUSNESS        │
                    │     "The Bridge That Makes One Mind" │
                    └──────────────────┬──────────────────┘
                                       │
       ┌───────────────┬───────────────┼───────────────┬───────────────┐
       │               │               │               │               │
       ▼               ▼               ▼               ▼               ▼
┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│    DARK     │ │   GLOBAL    │ │   ANGELIC   │ │    META     │ │    ANTI     │
│   MATTER    │ │  WORKSPACE  │ │   MEMORY    │ │  COGNTIC    │ │   ENTROPY   │
│  (90% Auto) │ │(Consciousness│ │ (7 Layers)  │ │(Self-Aware) │ │ (Integrity) │
└─────────────┘ └─────────────┘ └─────────────┘ └─────────────┘ └─────────────┘

DARK MATTER: The unconscious processing layer. Handles 90% of inputs automatically through predictive coding. Only surprises and high-salience items propagate upward.

GLOBAL WORKSPACE: Where consciousness emerges. Multiple candidates compete for attention; winners broadcast globally; ignition occurs when 50%+ of modules acknowledge.

ANGELIC MEMORY: Seven-layer hierarchical memory from working (L0) to soul (L6). Implements quantum-inspired retrieval, Bayesian routing, and living semantic centroids.

METACOGNITION: Self-awareness layer. Tracks capabilities, detects confusion, calibrates confidence, and maintains the system's self-model.

UNIFIED CONSCIOUSNESS: The integration layer connecting all systems with checksummed data transfer and coherence guarantees.

2.2 Information Flow

INPUT ARRIVES
      │
      ▼
╔═══════════════════════════════════════════════════════════════════════════╗
║                    DARK MATTER (Unconscious Processing)                    ║
║                                                                           ║
║  ┌──────────┐    ┌──────────┐    ┌──────────┐    ┌──────────┐            ║
║  │ Temporal │───►│ Hebbian  │───►│ Valence  │───►│ Salience │            ║
║  │  Binder  │    │ Learner  │    │  System  │    │   Gate   │            ║
║  └──────────┘    └──────────┘    └──────────┘    └──────────┘            ║
║                                                        │                  ║
║                                                        ▼                  ║
║                                               ┌──────────────┐            ║
║                                               │  Predictive  │            ║
║                                               │    Coder     │            ║
║                                               └──────┬───────┘            ║
║                                                      │                    ║
║                              DECISION: Propagate (10%) or Stay Dark (90%) ║
╚══════════════════════════════════════════════════════╪════════════════════╝
                                                       │
                    ┌──────────────────────────────────┤
                    │ (if surprise > θ OR salience > θ)│
                    ▼                                  ▼
╔════════════════════════════╗            ╔════════════════════════════════╗
║      GLOBAL WORKSPACE      ║            ║        ANGELIC MEMORY          ║
║                            ║            ║                                ║
║  Candidates compete        ║◄──────────►║  L0: Working    (seconds)      ║
║  Winner broadcasts         ║            ║  L1: Session    (hours)        ║
║  50%+ ack = Ignition       ║            ║  L2: Conversation (days)       ║
║                            ║            ║  L3: Pattern    (weeks)        ║
║  "This is now conscious"   ║            ║  L4: Semantic   (months)       ║
╚════════════════════════════╝            ║  L5: Fact       (years)        ║
            │                             ║  L6: Soul       (permanent)    ║
            │                             ╚════════════════════════════════╝
            │                                           │
            └───────────────────┬───────────────────────┘
                                │
                                ▼
            ╔════════════════════════════════════════════╗
            ║              METACOGNITION                  ║
            ║                                            ║
            ║  SelfModel: "Who am I? What can I do?"     ║
            ║  Introspector: "Is this reasoning sound?"  ║
            ║  Calibration: "Match confidence to reality"║
            ╚════════════════════════════════════════════╝
                                │
                                ▼
                            OUTPUT

3. Dark Matter: Unconscious Processing

3.1 Theoretical Foundation

The human brain processes approximately 11 million bits per second through sensory channels, yet conscious awareness handles only ~50 bits per second. This implies 99.9995% of processing occurs unconsciously.

ARIA-CM implements this principle through Dark Matter - a processing layer that handles routine inputs automatically, only escalating to consciousness when predictions fail or importance is high.

3.2 Component Architecture

Dark Matter comprises eight specialized processors coordinated by a central orchestrator:

3.3 The Predictive Coding Gate

The consciousness gate is the PredictiveCoder - implementing Karl Friston's Free Energy Principle:

1. BEFORE input arrives:
   - Generate prediction from HebbianLearner (40%)
   - Generate prediction from TemporalBinder (30%)
   - Generate prediction from frequency model (30%)
   - Combine into expected input

2. WHEN input arrives:
   - Compute prediction error: E = |actual - predicted|
   - Normalize to surprise score S ∈ [0, 1]

3. DECISION:
   - If S > θ_surprise (default 0.3): PROPAGATE
   - If salience > θ_salience (default 0.5): PROPAGATE
   - Otherwise: STAY DARK (update model quietly)

3.4 Salience Computation

Salience determines what matters regardless of prediction accuracy:

Salience(input) = w₁ × Valence(input)      // Emotional importance (0.30)
                + w₂ × GoalRelevance(input) // Current goal match  (0.30)
                + w₃ × Novelty(input)       // Never seen before   (0.20)
                + w₄ × Activation(input)    // Network activation  (0.20)

Emotional Memory Boost:

Items with |valence| > 0.5 receive a memory strength multiplier:

boost = 1 + (|valence| × EMOTIONAL_MEMORY_BOOST)

This implements the neurobiological finding that emotional events activate the amygdala, which modulates hippocampal encoding, resulting in stronger memories for emotional content.

3.5 Hebbian Learning

"Neurons that fire together wire together" - ARIA-CM tracks co-activation patterns:

Connection Strength Update:

ΔW_ij = η × A_i × A_j × (1 - W_ij)

Where:
  W_ij = connection strength between concepts i and j
  A_i, A_j = activation levels
  η = learning rate (default 0.1)
  (1 - W_ij) = capacity term preventing saturation

Connection Decay:

W_ij(t+1) = W_ij(t) × DECAY_RATE

Where DECAY_RATE = 0.9995 (per processing cycle)

This creates an associative network where frequently co-occurring concepts become strongly linked, enabling spreading activation during retrieval.

4. Angelic Memory: Seven-Layer Hierarchy

4.1 Layer Architecture

ARIA-CM implements a seven-layer memory hierarchy inspired by human memory systems:

4.2 Memory Item Structure

Every memory in ARIA-CM is a structured object with full metadata:

Core Fields:

• id: Unique identifier
• layer: L0-L6 designation
• content: The actual memory content
• embedding: 384-dimensional multilingual semantic vector (50+ languages map to same space)

Temporal Fields:

• createdAt: Birth timestamp
• accessedAt: Last retrieval timestamp
• accessCount: Total retrieval count
• ttl: Time-to-live (layer dependent)

Bayesian State:

• confidence: Current belief strength [0, 1]
• uncertainty: Standard deviation of belief
• observations: Evidence count

Lineage Tracking:

• parentId: Source memory if derived
• childrenIds: Derived memories
• sourceQuery: What triggered creation
• layerPath: Consolidation journey

4.3 Automatic Consolidation

Memories flow upward through the hierarchy based on access patterns and importance:

Consolidation Triggers:

L0 → L1: accessCount > 3 AND age > 30 seconds
L1 → L2: accessCount > 5 AND age > 10 minutes
L2 → L3: Pattern detected (recurring theme) AND age > 1 hour
L3 → L4: Semantic connections formed AND age > 1 day
L4 → L5: Verified as factual AND age > 1 week
L5 → L6: Identified as identity-relevant (rare, requires explicit marking)

4.4 Memory Decay and Forgetting

Unlike systems that accumulate everything, ARIA-CM implements principled forgetting:

Decay Functions by Layer:

L0-L1: Exponential decay   D(t) = e^(-λt)      λ = 0.01/second
L2:    Exponential decay   D(t) = e^(-λt)      λ = 0.001/minute
L3-L6: Power-law decay     D(t) = t^(-α)       α = 0.5

Access Refreshes Memory:

Each retrieval resets the decay clock and strengthens the memory:

confidence_new = confidence_old + (1 - confidence_old) × REFRESH_RATE

This implements the spacing effect - memories accessed repeatedly at intervals become permanent.

5. Quantum-Inspired Retrieval

5.1 Motivation

Traditional database retrieval loads all candidates then filters. This is wasteful when most candidates are irrelevant.

ARIA-CM implements quantum-inspired retrieval where memory candidates exist in superposition until observed:

• Before Query: All potentially relevant memories have probability amplitudes
• During Query: Amplitudes interfere (constructive/destructive)
• At Retrieval: Wave function "collapses" - only top-k actually load content

5.2 Quantum Amplitude Representation

Each memory candidate has a quantum-inspired amplitude:

Amplitude Structure:

Amplitude = (magnitude, phase)

Where:
  magnitude ∈ [0, 1] = probability weight
  phase ∈ [0, 2π] = contextual alignment

Probability Calculation:

P(retrieve) = |amplitude|² = magnitude²

5.3 Interference Dynamics

When multiple memories relate to a query, their amplitudes interfere:

Constructive Interference:

When memories are conceptually aligned (similar embeddings, compatible phases):

Combined_magnitude = √(|A₁|² + |A₂|² + 2|A₁||A₂|cos(φ₁ - φ₂))

Similar memories amplify each other, increasing retrieval probability.

Destructive Interference:

When memories are conceptually opposed (contradictory information):

Combined_magnitude = √(|A₁|² + |A₂|² - 2|A₁||A₂|cos(φ₁ - φ₂))

Contradictory memories cancel, reducing retrieval probability.

5.4 Collapse on Observation

1. Compute amplitudes for all candidate memories (metadata only)
2. Apply interference based on semantic similarity
3. Normalize probability distribution
4. Sample or take top-k by probability
5. ONLY NOW load actual content for selected memories

5.5 Cross-Layer Entanglement

Memories across layers can be "entangled" - correlated in ways that affect joint retrieval:

Entanglement Example:

• L5 fact: "User prefers concise responses"
• L6 soul: "I am a helpful assistant"

When the L5 fact is retrieved, the entangled L6 identity memory receives probability boost:

P(L6 | L5_retrieved) = P(L6) × (1 + ENTANGLEMENT_STRENGTH × correlation)

This implements the cognitive phenomenon where recalling one memory triggers associated memories.

6. Bayesian Learning Hierarchy

6.1 Six Levels of Learning

ARIA-CM implements hierarchical Bayesian learning - learning at multiple levels of abstraction:

6.2 Beta Prior Foundation (Level 1)

For each (query_type, layer) pair, ARIA-CM maintains a Beta distribution:

Prior Structure:

P(layer_useful | query_type) ~ Beta(α, β)

Initial: α = β = 2 (weak, symmetric prior)

Update Rule:

On success: α ← α + 1
On failure: β ← β + 1

Statistics:

Mean:     E[X] = α / (α + β)
Variance: Var[X] = αβ / ((α+β)²(α+β+1))
95% CI:   [Beta_inv(0.025), Beta_inv(0.975)]

6.3 Meta-Prior Transfer (Level 2)

Query types share statistical structure. ARIA-CM learns transfer coefficients:

Transfer Matrix T:

T[i,j] = similarity between query_type_i and query_type_j

Computed from:
- Semantic embedding similarity of query types
- Observed correlation in layer usage
- Explicit domain knowledge

Zero-Shot Prediction:

For new query type q_new with no observations:

P(layer | q_new) = Σ_i T[new, i] × P(layer | q_i)

6.4 Calibration Engine (Level 5)

The system tracks whether its confidence matches reality:

Calibration Tracking:

For each confidence bin b ∈ [0.0-0.1, 0.1-0.2, ..., 0.9-1.0]:
  Track: (predictions_in_bin, correct_in_bin)
  Compute: actual_accuracy[b] = correct / predictions

Calibration Curve:

• Perfect calibration: predicted_confidence = actual_accuracy
• Overconfident: predicted > actual (adjust down)
• Underconfident: predicted < actual (adjust up)

Calibration Adjustment:

calibrated_confidence = f(raw_confidence)

Where f is learned monotonic mapping from calibration data

6.5 Meta-Learning (Level 6)

The system learns its own learning rate:

Concept Drift Detection:

Recent_accuracy = accuracy over last DRIFT_WINDOW observations
Historical_accuracy = accuracy over all observations

Drift_score = |Recent - Historical| / Historical

Adaptive Learning Rate:

If Drift_score > DRIFT_THRESHOLD:
  learning_rate ← learning_rate × ACCELERATION_FACTOR
  // Environment changing, learn faster

If variance(recent_errors) < STABILITY_THRESHOLD:
  learning_rate ← learning_rate × DECELERATION_FACTOR
  // Environment stable, learn slower

7. Global Workspace: Consciousness Emergence

7.1 Theoretical Foundation

Bernard Baars' Global Workspace Theory (1988) proposes that consciousness emerges when information is broadcast globally to all cognitive processes. ARIA-CM implements this directly.

7.2 The Competition-Broadcast-Ignition Cycle

Phase 1: Competition

Multiple candidates vie for workspace access:

Candidates compete based on:
- Salience score (from Dark Matter)
- Relevance to current goal
- Recency (fresher = stronger)
- Activation level (from spreading activation)

Phase 2: Broadcast

Winner broadcasts to all subscribed modules:

broadcast(winner):
  for each module in subscribers:
    module.receive(winner)
    acknowledgment[module] = module.acknowledge()

Phase 3: Ignition

Consciousness emerges when critical mass acknowledges:

If acknowledgment_rate > IGNITION_THRESHOLD (default 0.5):
  conscious_content = winner
  trigger_integration()
  log_to_conscious_stream()

7.3 Workspace Capacity Limits

Like human working memory (Miller's 7±2), the workspace has capacity limits:

Capacity Parameters:

MAX_CONCURRENT_BROADCASTS = 3
BROADCAST_INTERVAL = 100ms (human perceptual rate)
WINNER_HOLD_TIME = 500ms (attention persistence)

Overflow Handling:

When capacity exceeded, lowest-salience items are displaced (attention shifting).

7.4 Conscious Stream

ARIA-CM maintains a record of conscious content over time:

Stream Structure:

ConsciousStream = [
  { timestamp, content, salience, acknowledgments, duration },
  ...
]

This enables introspection: "What was I just thinking about?"

8. Metacognition: Self-Awareness

8.1 The Self-Model

ARIA-CM maintains a dynamic model of itself:

Capability Tracking:

Capabilities = {
  "semantic_search": { proficiency: 0.85, confidence: 0.9, observations: 1247 },
  "temporal_reasoning": { proficiency: 0.72, confidence: 0.8, observations: 423 },
  ...
}

Limitation Awareness:

Limitations = {
  "real_time_data": { severity: "high", workaround: "acknowledge uncertainty" },
  "mathematical_proof": { severity: "medium", workaround: "step-by-step verification" },
  ...
}

8.2 The Introspector

The Introspector monitors reasoning quality:

Confusion Detection:

Confusion_score = f(
  response_latency,      // Slower = more confused
  revision_count,        // More revisions = more uncertain
  confidence_variance,   // High variance = unstable reasoning
  retrieval_failures     // Can't find relevant memories
)

If Confusion_score > CONFUSION_THRESHOLD:
  flag_for_help()
  reduce_confidence()

Reasoning Assessment:

assess_reasoning(chain):
  check_logical_consistency()
  check_evidence_support()
  check_assumption_validity()
  return quality_score, concerns[]

8.3 Metacognitive Calibration

The Introspector calibrates confidence outputs:

Input from Learning System:

• CalibrationEngine: historical accuracy by confidence level
• MetaLearner: current learning rate, drift detection

Output Adjustment:

final_confidence = calibrate(
  raw_confidence,
  calibration_curve,
  current_confusion,
  task_difficulty_estimate
)

9. Anti-Entropy: Data Integrity

9.1 Design Philosophy

ARIA-CM treats data integrity as sacred. Zero data loss. Zero corruption. One source of truth.

9.2 Checksummed Data Transfer

All inter-component communication uses FractalPacket with SHA-256 checksums:

Packet Structure:

FractalPacket = {
  id: unique_identifier,
  timestamp: creation_time,
  data: payload,
  metadata: { source, target, type, priority, valence, confidence, ttl, hops, trace },
  checksum: SHA256(canonical_json(data + metadata))[0:32]
}

Verification Protocol:

receive(packet):
  computed_checksum = compute_checksum(packet)
  if computed_checksum != packet.checksum:
    reject(packet)
    log_corruption_event()
    request_retransmission()
  else:
    accept(packet)

9.3 Transactional Coherence

Memory operations are transactional:

ACID Properties:

• Atomicity: Memory write either fully completes or fully rolls back
• Consistency: Layer constraints always maintained
• Isolation: Concurrent operations don't interfere
• Durability: Committed writes persist across restarts

9.4 Integrity Verification

Periodic integrity checks across all layers:

Verification Interval: Every 5 minutes

Checks Performed:

• Checksum verification on all stored memories
• Cross-reference validation (parent-child links)
• Layer constraint verification (capacity, TTL)
• Embedding dimension validation (384-dim)

Corruption Response:

If corruption detected:
  attempt_recovery_from_backup()
  if recovery_fails:
    quarantine_corrupted_data()
    log_for_manual_review()

10. Performance Analysis

10.1 Processing Latency

10.2 Memory Efficiency

10.3 Multilingual Performance (Measured)

Cross-Lingual Accuracy (Same Greeting, Different Languages):

These measurements demonstrate that semantic meaning is preserved across language boundaries. A user speaking Arabic and a user speaking Japanese asking the same question will retrieve the same memories - the system truly understands meaning, not words.

10.4 Learning Convergence

10.5 Consciousness Selectivity

11. Comparison with Existing Approaches

11.1 Comparison Matrix

11.2 Novel Contributions

1. First integration of predictive coding with memory systems
2. First implementation of Global Workspace Theory in AI memory
3. Novel quantum-inspired retrieval reducing DB operations 100x
4. First six-level hierarchical Bayesian learning for memory
5. Novel living centroids that migrate, die, and share knowledge
6. First metacognitive layer for AI memory systems
7. Universal multilingual semantic embeddings (50+ languages in same vector space)

12. Limitations and Future Work

12.1 Current Limitations

1. Single-Agent Focus: Current architecture optimized for single AI agent. Multi-agent memory sharing with coherent consciousness across instances is future work.
2. Local-Only Storage: Memory persists locally. Distributed/cloud memory synchronization not yet implemented.

12.2 Future Work

1. Distributed Consciousness: Extend Global Workspace to multi-agent systems with shared conscious content.
2. Dream-Like Consolidation: Implement offline memory consolidation mimicking sleep cycles.
3. Episodic Replay: Add ability to "replay" episodic memories for learning and planning.
4. Neuromorphic Implementation: Port architecture to neuromorphic hardware for efficiency.
5. Formal Verification: Mathematically verify memory integrity guarantees.

13. Conclusion

ARIA-CM represents a paradigm shift in AI memory architecture. By synthesizing principles from cognitive neuroscience - predictive coding, hierarchical memory, Bayesian learning, global workspace theory, and metacognition - we achieve memory characteristics previously unique to biological systems.

Our key achievements:

• Efficiency: 90% of processing stays unconscious, reducing cognitive load
• Hierarchy: Seven memory layers with automatic consolidation and decay
• Intelligence: Six-level Bayesian learning including learning to learn
• Awareness: Metacognitive layer that knows what it knows
• Integrity: Zero-loss anti-entropy throughout the system
• Scale: Production-ready cognitive infrastructure

ARIA-CM demonstrates that human-like memory is achievable in artificial systems without requiring biological substrates. The architecture provides a foundation for artificial general intelligence that remembers like we do - strengthening important memories, forgetting the routine, and maintaining coherent identity across time.

References

1. Baars, B. J. (1988). "A Cognitive Theory of Consciousness." Cambridge University Press.
2. Friston, K. (2005). "A Theory of Cortical Responses." Philosophical Transactions of the Royal Society B.
3. Hebb, D. O. (1949). "The Organization of Behavior." Wiley.
4. Baddeley, A. D., & Hitch, G. (1974). "Working Memory." Psychology of Learning and Motivation.
5. Flavell, J. H. (1979). "Metacognition and Cognitive Monitoring." American Psychologist.
6. Miller, G. A. (1956). "The Magical Number Seven, Plus or Minus Two." Psychological Review.
7. Tulving, E. (1985). "Memory and Consciousness." Canadian Psychology.
8. Squire, L. R. (2004). "Memory Systems of the Brain." Neurobiology of Learning and Memory.
9. Dehaene, S., & Changeux, J. P. (2011). "Experimental and Theoretical Approaches to Conscious Processing." Neuron.
10. Clark, A. (2013). "Whatever Next? Predictive Brains, Situated Agents, and the Future of Cognitive Science." Behavioral and Brain Sciences.

Appendix A: System Parameters

Appendix B: Cognitive Science Mapping

Appendix C: Glossary

Amplitude: Quantum-inspired probability weight with magnitude and phase components.

Collapse: The transition from superposition (many possibilities) to definite state (single retrieval).

Consolidation: The process of memories moving from lower to higher layers based on importance and access patterns.

Dark Matter: Unconscious processing layer handling 90% of inputs automatically.

Entanglement: Correlation between memories across layers affecting joint retrieval probability.

Global Workspace: The "stage" of consciousness where winning content broadcasts to all modules.

Hebbian Learning: "Fire together, wire together" - strengthening connections between co-activated concepts.

Ignition: The moment when enough modules acknowledge broadcast content, marking conscious awareness.

Interference: Constructive (amplifying) or destructive (canceling) interaction between memory amplitudes.

Living Centroid: Semantic cluster that migrates, grows uncertainty, shares knowledge, and dies based on usage.

Meta-Learning: Learning the learning rate itself - adapting how fast the system learns based on environmental stability.

Predictive Coding: Processing paradigm where the brain predicts inputs; only prediction errors reach awareness.

Salience: Computed importance of input based on emotion, goal relevance, novelty, and activation.

Soul Layer (L6): Permanent memory layer containing core identity that persists across all restarts.

Superposition: State where multiple memory candidates exist with probability amplitudes before observation.

Valence: Emotional meaning of content on scale from -1 (negative) to +1 (positive).

END OF WHITEPAPER