---
This dissertation argues that computational neuroscience offers a principled—and actionable—blueprint for building AI assistants that operate continuously rather than responding to discrete queries. The core insight is not that brains are perfect systems to copy, but that 500 million years of evolution have discovered solutions to problems we're now encountering in AI: maintaining coherent state across time, learning without catastrophic forgetting, allocating attention across competing demands, and acting under uncertainty while remaining responsive to the unexpected.
Contemporary AI assistants are fundamentally reactive—they receive prompts, generate responses, and forget. An always-on assistant requires something categorically different: persistent state that evolves rather than resets, learning that compounds rather than overwrites, attention that allocates rather than focuses-or-ignores, and action selection that balances exploitation with exploration across timescales from milliseconds to months.
This curriculum's six units—dynamics, coding, plasticity, circuits, predictive processing, and large-scale integration—are not independent topics but interlocking pieces of a unified architecture. The membrane equation teaches us about state accumulation and temporal filtering. Neural coding reveals the tradeoffs between efficiency and robustness. Plasticity mechanisms show how to learn continuously without destroying prior knowledge. Circuit motifs provide building blocks for routing, gating, and competition. Predictive processing offers a normative framework for integrating uncertain information. And large-scale integration shows how specialized modules coordinate into coherent behavior.
The thesis: An effective always-on assistant should be organized as a hierarchical generative model with multi-timescale dynamics, complementary memory systems, precision-weighted inference, and neuromodulatory control over the exploration-exploitation tradeoff. This is not a metaphor—it is a concrete architectural specification.
The bet: That the brain's solutions to continuous cognition generalize beyond biological constraints, and that implementing these mechanisms in silicon will yield assistants that are more coherent, more adaptive, and more reliable than current architectures.
What follows: An architectural blueprint, a mechanism integration guide, an honest assessment of assumptions and failure modes, and a phased implementation plan. This is the capstone—the synthesis that makes the curriculum actionable.
---
Current AI assistants are stateless pattern-matchers wrapped in conversation scaffolding. Each interaction is essentially independent—prior context is injected as text, but there's no genuine persistence of state, no differential treatment of recent versus ancient history, no accumulated model of the user that evolves through interaction.
The brain solves a harder version of this problem. A biological organism must:
1. Maintain coherent identity across waking hours, sleep cycles, and years
2. Learn continuously from experience without forgetting critical skills
3. Allocate finite resources (attention, energy, processing) across competing demands
4. Act under uncertainty while remaining responsive to genuine surprises
5. Coordinate multiple specialized systems into unified behavior
An always-on assistant faces the same challenges. It cannot afford to restart from scratch each session. It cannot treat all information as equally weighted. It cannot persist indefinitely on a single task while the world changes. It cannot ignore its own uncertainty.
The brain's solution is not a single algorithm but an organized architecture—a set of interacting mechanisms that solve different subproblems while remaining coordinated. The six units of this curriculum map to six distinct architectural requirements:
| Unit | Core Mechanism | Architectural Requirement |
|------|----------------|--------------------------|
| 0: Dynamics | Membrane integration, attractors, timescales | State evolution that filters noise while remaining responsive |
| 1: Coding | Rate vs. temporal, sparse vs. distributed | Representation formats that trade off efficiency against robustness |
| 2: Plasticity | Hebbian learning, STDP, homeostasis | Learning rules that compound knowledge without catastrophic forgetting |
| 3: Circuits | Gating, WTA, normalization | Routing and competition that allocate resources dynamically |
| 4: Predictive Processing | Free energy, precision weighting | Inference under uncertainty with coherent belief maintenance |
| 5: Integration | Thalamus, neuromodulation, networks | Coordination across specialized modules into unified behavior |
The insight is that these mechanisms are interdependent. You cannot have stable learning (Unit 2) without homeostatic dynamics (Unit 0). You cannot have coherent inference (Unit 4) without proper circuit gating (Unit 3). You cannot have unified behavior (Unit 5) without precision-weighted coding (Unit 1).
Translating these insights to AI means:
1. State is not just context—it's evolved. The assistant maintains internal states that accumulate evidence, decay stale beliefs, and transition between discrete modes. This is the membrane equation applied to cognitive state.
2. Representations are not uniform—they're stratified. Different information requires different coding schemes: sparse codes for discrete categories (which brain should handle this?), distributed codes for similarity relationships (what does this remind me of?), temporal codes where timing matters (when should I interrupt?).
3. Learning is not fine-tuning—it's consolidated. New information enters fast episodic stores before selective consolidation into slower parametric memory. Importance weights protect critical knowledge from overwriting.
4. Attention is not binary—it's competitive. Multiple processes compete for computational resources through soft winner-take-all dynamics. Normalization ensures coherent allocation across the system.
5. Uncertainty is not ignored—it's represented. Every belief carries precision metadata. Inference weights new evidence against priors based on estimated reliability. Actions are selected to minimize expected uncertainty, not just immediate error.
6. Control is not centralized—it's modulated. Instead of a single decision-maker, neuromodulatory signals adjust parameters across the system: exploration/exploitation balance, learning rate, attention breadth.
---
┌─────────────────────────────────────────────────────────────────────────────┐
│ META-COGNITIVE LAYER │
│ ┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐ │
│ │ Uncertainty │ │ Mode Control │ │ Health Monitor │ │
│ │ Estimation │ │ (Neuromod) │ │ (Failure Detect)│ │
│ └────────┬────────┘ └────────┬────────┘ └────────┬────────┘ │
│ │ Global Modulation │ │ │
│ ▼ ▼ ▼ │
├─────────────────────────────────────────────────────────────────────────────┤
│ HIERARCHICAL GENERATIVE MODEL │
│ ┌─────────────────────────────────────────────────────────────────────┐ │
│ │ Level 3: Goals & Intents (hours-days) │ │
│ │ - What does the user want? │ │
│ │ - What should I be doing? │ │
│ └──────────────────────────┬──────────────────────────────────────────┘ │
│ │ predictions ↓↑ errors │
│ ┌──────────────────────────▼──────────────────────────────────────────┐ │
│ │ Level 2: Contexts & Tasks (minutes-hours) │ │
│ │ - What conversation is this? │ │
│ │ - What's the current task? │ │
│ └──────────────────────────┬──────────────────────────────────────────┘ │
│ │ predictions ↓↑ errors │
│ ┌──────────────────────────▼──────────────────────────────────────────┐ │
│ │ Level 1: Actions & Utterances (seconds-minutes) │ │
│ │ - What should I say now? │ │
│ │ - What action should I take? │ │
│ └──────────────────────────┬──────────────────────────────────────────┘ │
│ │ predictions ↓↑ errors │
│ ┌──────────────────────────▼──────────────────────────────────────────┐ │
│ │ Level 0: Observations (sub-second) │ │
│ │ - Raw inputs from all channels │ │
│ └─────────────────────────────────────────────────────────────────────┘ │
├─────────────────────────────────────────────────────────────────────────────┤
│ COMPLEMENTARY MEMORY SYSTEMS │
│ ┌─────────────────────────┐ ┌─────────────────────────────────────┐ │
│ │ Episodic Store │ │ Parametric Knowledge │ │
│ │ (Vector DB, fast write) │◄───►│ (LLM weights, slow update) │ │
│ │ - Specific events │ Con │ - General patterns │ │
│ │ - Recent interactions │ sol │ - Learned associations │ │
│ │ - Tagged for importance │ ida │ - Importance-weighted │ │
│ └─────────────────────────┘ tion└─────────────────────────────────────┘ │
├─────────────────────────────────────────────────────────────────────────────┤
│ PRECISION ESTIMATION │
│ ┌─────────────────────────────────────────────────────────────────────┐ │
│ │ Per-source: calendar > inferred_schedule > guessed_intent │ │
│ │ Contextual: high precision when signal strong, low when ambiguous │ │
│ │ Bounds: min ≤ precision ≤ max (prevent collapse) │ │
│ │ Health: monitor for drift, trigger recalibration │ │
│ └─────────────────────────────────────────────────────────────────────┘ │
├─────────────────────────────────────────────────────────────────────────────┤
│ ACTIVE INFERENCE ENGINE │
│ ┌─────────────────────────────────────────────────────────────────────┐ │
│ │ Policy Selection: minimize expected free energy │ │
│ │ - Epistemic value: actions that reduce uncertainty │ │
│ │ - Pragmatic value: actions that achieve goals │ │
│ │ - Temporal horizon: extended planning, not greedy │ │
│ └─────────────────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────────────────┘
The core of the system is a hierarchical predictive model with four levels operating at distinct timescales:
Level 0 (Sub-second): Raw sensory processing. Encodes incoming messages, sensor data, and events. Operates in feedforward mode for speed. Computes prediction errors against Level 1 expectations.
Level 1 (Seconds-minutes): Action-level processing. Predicts next likely observations. Selects responses. Maintains working memory for immediate context. Updates via precision-weighted prediction errors from Level 0.
Level 2 (Minutes-hours): Task and context level. Tracks which conversation we're in, what task we're doing, what the user seems to want. Updates more slowly—requires accumulated evidence before revising.
Level 3 (Hours-days): Goal and intent level. Models user's longer-term goals, relationship dynamics, recurring patterns. Protects against overwriting by importance weights. Updates during consolidation cycles.
Critical design decisions:
Inspired by hippocampal-neocortical complementary learning systems:
Episodic Store:
Parametric Knowledge:
Consolidation Pathway:
Maintains confidence estimates for all information sources:
Source hierarchy (default):
explicit_user_statement > calendar_data > email_content >
inferred_intent > behavioral_guess > default_prior
Dynamic adjustment:
Bounds and health:
Instead of fixed parameters, a meta-controller adjusts system-wide settings:
Exploration-Exploitation (LC analog):
Attention-Learning Coupling (ACh analog):
Arousal Level (NE analog):
Monitors the system itself:
Uncertainty Estimation:
Health Monitoring:
Mode Control:
The system uses canonical circuit motifs for specific functions:
| Function | Motif | Implementation |
|----------|-------|----------------|
| Input filtering | Feedforward inhibition | Time-windowed attention, urgency gating |
| Context maintenance | Recurrent excitation + gating | LSTM-like memory cells with input/forget gates |
| Response selection | Winner-take-all | Softmax with adaptive temperature |
| Resource allocation | Divisive normalization | Attention weights normalized across sources |
| Brain routing | Soft WTA + threshold | Confidence-weighted routing with fallback |
| Sequence execution | Discrete attractors + synfire | Step states with explicit transitions |
| Mode switching | Competitive inhibition | Internal vs. external processing networks |
---
The previous section described components. This section describes how they interact—because that's where most implementations fail. A system of well-designed components can still fail catastrophically if the interactions are wrong.
The flow:
1. Observation arrives (user message, sensor data, event)
2. Level 0 encodes observation into standard representation
3. Prediction error computed: ε = observation - prediction from Level 1
4. Precision weighting applied: ξ = Π · ε
5. Belief update propagates upward, each level updating based on weighted errors from below
6. Updated beliefs generate new predictions flowing downward
7. If prediction error exceeds threshold at any level, trigger full re-analysis (not just incremental update)
Critical interaction: Precision weights are computed by a separate module but applied within the hierarchical model. The precision module needs access to both observations and predictions to estimate reliability.
Episodic store interaction with hierarchy:
Consolidation timing:
Critical interaction: Consolidation cannot run during active interaction—the system must detect appropriate windows. Neuromodulatory state (low arousal) gates consolidation permission.
Active inference loop:
1. Generate candidate actions (based on current beliefs and policy prior)
2. For each candidate, predict outcomes (forward model)
3. Compute expected free energy: uncertainty about outcomes (epistemic) + distance from goals (pragmatic)
4. Select action that minimizes expected free energy
5. Execute action, observe outcome
6. Update forward model based on prediction error
Critical interaction: Action selection depends on precision estimates. If precision is low, epistemic value dominates (actions that gather information). If precision is high, pragmatic value dominates (actions that achieve goals). This is how the system balances asking questions versus acting.
Mode control effects:
Critical interaction: Mode control modulates nearly every other component. This is intentional—neuromodulation is the brain's mechanism for system-wide coordination. But it means mode control is a single point of failure. Health monitoring must watch for pathological mode states.
Everything described above requires temporal coordination. The system operates in nested loops:
Fast loop (sub-second):
Medium loop (seconds):
Slow loop (minutes):
Offline loop (hours, when inactive):
---
This architecture makes several bets that may prove wrong:
Bet 1: Hierarchy generalizes
The brain's hierarchical organization evolved for biological constraints (wiring costs, metabolic limits, physical embodiment). We assume the computational benefits (efficient inference, temporal separation, modular specialization) transfer to silicon. This may be true for the same reasons it's true for biological systems—complexity management—but we lack proof.
Bet 2: Predictive processing is normatively correct
We assume the brain's predictive architecture reflects something close to optimal inference under uncertainty, not just evolutionary path dependence. The free energy framework is mathematically elegant but empirically contested. Alternative frameworks (attention schema theory, global workspace theory, integrated information theory) make different predictions.
Bet 3: Separation of timescales helps
We assume that enforcing distinct timescales for different hierarchical levels improves stability and coherence. This is supported by dynamical systems theory but may introduce problems (temporal misalignment, delays in propagating critical updates).
Bet 4: Precision estimation is tractable
We assume we can reliably estimate the reliability of different information sources. In practice, precision is hard to estimate—especially for rare events, adversarial inputs, or distribution shift. Bad precision estimates could be worse than no precision weighting at all.
Bet 5: Neuromodulatory control scales
We assume that system-wide modulation via a small number of control signals (exploration, arousal, learning rate) provides sufficient coordination. This may work at small scale but fail when the system has thousands of specialized modules.
Computational cost:
This architecture is more expensive than a stateless prompt-response model. Maintaining hierarchical state, computing precision weights, running consolidation cycles—all cost compute. The bet is that improved performance justifies the cost, but this is unproven.
Complexity:
More mechanisms mean more ways to fail. Each interaction point is a potential bug. Debugging will be harder than simpler architectures.
Interpretability:
Hierarchical generative models with precision weighting and neuromodulatory control are harder to interpret than feedforward networks. Understanding why the system made a particular decision requires understanding multiple interacting mechanisms.
Evaluation:
Standard benchmarks (task accuracy, response quality) don't capture what this architecture is designed for. We need new metrics: coherence over time, learning without forgetting, appropriate confidence calibration, exploration/exploitation balance.
Biological fidelity:
We're inspired by neuroscience but not constrained by it. Some biological mechanisms may be irrelevant (e.g., metabolic constraints) while others may be critical but unimplemented (e.g., detailed neuromodulator receptor dynamics). We don't know which is which.
Consciousness and phenomenology: This is a functional architecture, not a theory of mind. Whether it "experiences" anything is outside scope.
Social modeling: The architecture focuses on individual cognition. Modeling other agents (theory of mind, social dynamics) is orthogonal.
Embodiment: We assume inputs arrive as messages and outputs are actions/responses. Physical embodiment would change the architecture significantly.
Safety and alignment: The architecture provides mechanisms (precision weighting, uncertainty estimation) that could support safe behavior, but alignment is not the focus.
---
These would render the system unusable:
Precision collapse:
Runaway consolidation:
Hierarchical desynchronization:
Mode lock:
These reduce quality but don't render the system unusable:
Prior overfitting (hallucination):
Catastrophic forgetting:
Exploration paralysis:
Context fragmentation:
These arise from mechanism interactions:
Precision-consolidation conflict:
Mode-hierarchy misalignment:
Memory-prediction conflict:
---
Goal: Establish infrastructure and baselines.
Deliverables:
Dependencies: None (starting point)
Risks: Architecture may need revision as mechanisms are added.
Success criteria: System operates with quality comparable to non-hierarchical baseline.
Goal: Implement and validate individual mechanisms.
Subphases:
1a: Predictive Hierarchy (Weeks 5-6)
1b: Precision Weighting (Weeks 7-8)
1c: Memory Systems (Weeks 9-10)
1d: Circuit Motifs (Weeks 11-12)
Dependencies: Phase 0 complete.
Risks: Mechanisms may interact unexpectedly; budget time for debugging.
Success criteria: Each mechanism shows significant improvement over ablated version on targeted metrics.
Goal: Connect mechanisms into coherent system.
Subphases:
2a: Hierarchy-Memory Integration (Weeks 13-14)
2b: Precision-Inference Integration (Weeks 15-16)
2c: Neuromodulatory Control (Weeks 17-18)
Dependencies: Phase 1 complete.
Risks: Integration bugs; mode control may be difficult to tune.
Success criteria: Integrated system outperforms sum-of-parts (synergy detected).
Goal: Detect and handle failure modes.
Subphases:
3a: Failure Detection (Weeks 19-20)
3b: Recovery Mechanisms (Weeks 21-22)
3c: Stress Testing (Weeks 23-24)
Dependencies: Phase 2 complete.
Risks: Failure modes may interact in unexpected ways.
Success criteria: System operates stably for extended periods under varied conditions.
Phase 0: Foundation
│
▼
Phase 1: Core Mechanisms
│
├── 1a: Predictive Hierarchy ──┐
├── 1b: Precision Weighting ───┤
├── 1c: Memory Systems ────────┼── (can run in parallel)
└── 1d: Circuit Motifs ────────┘
│
▼
Phase 2: Integration
│
├── 2a: Hierarchy-Memory ───┐
├── 2b: Precision-Inference ┼── (sequential)
└── 2c: Neuromodulatory ────┘
│
▼
Phase 3: Robustness
│
├── 3a: Failure Detection ─┐
├── 3b: Recovery ──────────┼── (sequential)
└── 3c: Stress Testing ────┘
---
After 3000+ words on mechanisms, circuits, and failure modes, what's the one thing worth remembering?
The brain doesn't process information—it maintains a world model and minimizes surprise.
This is the fundamental reframe. Current AI assistants are stimulus-response machines: input → process → output → forget. The brain—and an effective always-on assistant—is different: it maintains an evolving model of the world, predicts what should happen, notices when predictions fail, and updates beliefs accordingly.
This isn't just a different architecture. It's a different goal. The system's purpose isn't to answer queries—it's to reduce uncertainty about the user, the context, and what should happen next. Answering queries is a byproduct.
From this single insight, everything else follows:
The mechanisms in this curriculum aren't arbitrary biological details. They're solutions to the core problem of maintaining coherent beliefs in a changing world. The brain discovered these solutions through evolution. We can implement them through engineering.
The bet is that this will work—that the brain's solutions generalize beyond biological constraints, that silicon can run these algorithms faster and more reliably than neurons, that an AI assistant built on these principles will be more coherent, more adaptive, and more useful than one built on prompt-response architecture.
It's a bet worth making.
---
Foundational neuroscience:
Predictive processing:
Continual learning:
---
Dissertation completed: 2026-02-14
Curriculum: Computational Neuroscience Foundations
Word count: ~4,200