Variational Free-Energy Guided Recall: How Active Inference Resolves Situational Ambiguity in AI Memory
Executive Summary: The Vector Distance Paradox
In modern Retrieval-Augmented Generation (RAG) and autonomous agent architectures, memory retrieval is almost universally treated as a static geometric search problem. A query string is transformed into an embedding vector, scanned against an approximate nearest neighbor (ANN) index using cosine similarity or Euclidean distance, and the top-$K$ nearest vectors are dumped into the prompt context window.
While computationally convenient, this paradigm suffers from a profound cognitive limitation known as the Vector Distance Paradox:
The Vector Distance Paradox: A memory can have near-perfect semantic cosine similarity to a prompt while contributing zero epistemic value toward resolving the agent's current uncertainty—or worse, amplifying confusion by reinforcing irrelevant surface associations.
In biological cognition, memory recall is not a passive lookup table. The brain is an active inference engine (Friston, 2010). It maintains an internal generative model of the self and the world, continuously seeking memories that minimize variational free energy—the mathematical upper bound on surprise and epistemic ambiguity.
In this deep architectural exploration, we unpack how Spector Memory introduces Free-Energy Guided Recall (FEGR): a neurocomputational retrieval framework that models the agent's mental state as a continuous Gaussian posterior distribution, evaluates candidate memories by their mathematical capacity to reduce variational free energy ($\Delta \mathcal{F}$), and accelerates precision-weighted Bayesian inference using hardware-accelerated SIMD vector pipelines.
flowchart LR
subgraph Traditional ["Traditional Vector Search"]
Q1["Query Vector"] --> CS["Cosine Distance"]
CS --> TopK1["Top-K By Semantic Similarity\n(High Epistemic Redundancy)"]
end
subgraph ActiveInference ["Active Inference Guided Recall"]
Q2["Sensory Observation o(t)"] --> GSM["Generative Prior p(s|m)\n+ Mental State q(s_t)"]
GSM --> FERS["Free-Energy Relevance Scoring\nΔF Minimization + Precision"]
FERS --> TopK2["Optimal Free-Energy Reducers\n(Zero Ambiguity, Maximum Grounding)"]
end
1. Theoretical Foundations: Active Inference & Free Energy
1.1 The Free Energy Principle
The Free Energy Principle, formulated by neuroscientist Karl Friston, states that any self-organizing system that resists decay and maintains its structural integrity must minimize an information-theoretic quantity called Variational Free Energy ($\mathcal{F}$).
Let:
- $\mathbf{s} \in \mathbb{R}^D$ represent the hidden, unobserved internal cognitive state of the agent.
- $\mathbf{o} \in \mathbb{R}^D$ represent the incoming sensory observation (e.g., user query, environment payload).
- $\mathbf{m}$ represent the agent's generative self-model (identity priors, values, persona definition).
The generative model defines a joint probability distribution over observations and hidden states:
$$p(\mathbf{o}, \mathbf{s} \mid \mathbf{m}) = p(\mathbf{o} \mid \mathbf{s}) \cdot p(\mathbf{s} \mid \mathbf{m})$$
Because computing the true posterior $p(\mathbf{s} \mid \mathbf{o}, \mathbf{m})$ directly is intractable in high-dimensional continuous state spaces, the agent maintains an approximate variational posterior distribution $q(\mathbf{s})$.
Variational Free Energy $\mathcal{F}(q, \mathbf{o})$ is defined as:
$$\mathcal{F}(q, \mathbf{o}) = \underbrace{D_{\text{KL}}\big[q(\mathbf{s}) \,\|\, p(\mathbf{s} \mid \mathbf{m})\big]}_{\text{Cognitive Complexity / Divergence}} - \underbrace{\mathbb{E}_{q(\mathbf{s})}\big[\ln p(\mathbf{o} \mid \mathbf{s})\big]}_{\text{Observation Accuracy / Likelihood}}$$
flowchart TD
VFE["Variational Free Energy: F(q, o)"]
subgraph Complexity ["1. Cognitive Complexity Penalty (D_KL)"]
KL["KL Divergence: D_KL( q(s) || p(s|m) )"]
KL_1["Penalizes beliefs drifting from core identity"]
KL_2["Prevents persona and value fragmentation"]
KL --> KL_1 & KL_2
end
subgraph Accuracy ["2. Sensory Accuracy Term (Likelihood)"]
LL["Expected Log-Likelihood: E_q( ln p(o|s) )"]
LL_1["Measures explanatory grounding on observations"]
LL_2["Rewards accurate situational context"]
LL --> LL_1 & LL_2
end
VFE --> Complexity
VFE --> Accuracy
- Complexity ($D_{\text{KL}}$): Penalizes beliefs that diverge wildly from the agent's foundational identity and prior expectations.
- Accuracy ($\mathbb{E}_q[\ln p(\mathbf{o} \mid \mathbf{s})]$): Measures how well the internal state explains the current observation.
Minimizing free energy simultaneously maximizes accuracy while keeping cognitive complexity bounded.
1.2 The Gaussian Parametric Parameterization
In Spector Memory, hidden states and priors are parameterized as multivariate Gaussian distributions with diagonal precision matrices (inverse variances) to enable closed-form, deterministic microsecond execution:
$$\begin{aligned} q(\mathbf{s}) &= \mathcal{N}\left(\boldsymbol{\mu}_q, \text{diag}(\boldsymbol{\pi}_q^{-1})\right) \\ p(\mathbf{s} \mid \mathbf{m}) &= \mathcal{N}\left(\boldsymbol{\mu}_p, \text{diag}(\boldsymbol{\pi}_p^{-1})\right) \end{aligned}$$
where:
- $\boldsymbol{\mu} \in \mathbb{R}^D$ is the expected mean state vector.
- $\boldsymbol{\pi} \in \mathbb{R}^D$ is the precision vector, where $\pi_d = \frac{1}{\sigma_d^2} > 0$. High precision signifies absolute confidence along a cognitive dimension; low precision reflects uncertainty.
Under this parameterization, the Kullback-Leibler divergence has an exact closed-form solution:
$$D_{\text{KL}}\big[q(\mathbf{s}) \,\|\, p(\mathbf{s} \mid \mathbf{m})\big] = \frac{1}{2} \sum_{d=1}^D \left[ \frac{\pi_{p,d}}{\pi_{q,d}} + \pi_{p,d} (\mu_{q,d} - \mu_{p,d})^2 - 1 - \ln\left(\frac{\pi_{p,d}}{\pi_{q,d}}\right) \right]$$
And assuming an observation mapping $p(\mathbf{o} \mid \mathbf{s}) = \mathcal{N}(\mathbf{s}, \text{diag}(\boldsymbol{\pi}_o^{-1}))$, the expected log-likelihood evaluates to:
$$\mathbb{E}_{q(\mathbf{s})}\big[\ln p(\mathbf{o} \mid \mathbf{s})\big] = -\frac{1}{2} \sum_{d=1}^D \left[ \ln(2\pi) - \ln(\pi_{o,d}) + \pi_{o,d} \left( (\mu_{q,d} - o_d)^2 + \frac{1}{\pi_{q,d}} \right) \right]$$
2. Memory Conditioning & $\Delta \mathcal{F}$ Ambiguity Reduction
2.1 The Epistemic Value of Recall
When an agent encounters a query $o(t)$, it experiences cognitive tension: its current posterior $q(s_t)$ has a non-zero free energy $\mathcal{F}_0 = \mathcal{F}(q(s_t), o(t))$.
Each candidate memory $\mu_i$ stored in Spector's 4-tier cortical storage acts as a conditioning evidence vector $e_{\mu_i}$ with memory confidence precision $\pi_{\mu_i}$.
sequenceDiagram
autonumber
participant S as Sensory Query o(t)
participant T as Mental State Tracker
participant F as Free Energy Calculator
participant R as Candidate Memory Pool
participant G as FreeEnergyGuidedRelay
S->>T: Query observation o(t)
T->>F: Compute baseline free energy F_0
R->>G: Stream candidate memories
loop For Each Candidate Memory
G->>F: Condition posterior q(s | mu_i) via Bayesian Fusion
F->>F: Compute conditional free energy F(q(s | mu_i), o)
F->>G: Delta_F = F_0 - F(q(s | mu_i))
end
G->>G: Compute Multi-Factor FERS Score
G-->>R: Rank candidates by ambiguity reduction
Conditioning the current belief state on memory $\mu_i$ yields an updated posterior $q(\mathbf{s} \mid \mu_i) = \mathcal{N}(\boldsymbol{\mu}_{q|\mu}, \text{diag}(\boldsymbol{\pi}_{q|\mu}^{-1}))$ via precision-weighted Bayesian cue combination:
$$\begin{aligned} \boldsymbol{\pi}_{q|\mu} &= \boldsymbol{\pi}_q + \boldsymbol{\pi}_{\mu_i} \\ \boldsymbol{\mu}_{q|\mu} &= \frac{\boldsymbol{\pi}_q \odot \boldsymbol{\mu}_q + \boldsymbol{\pi}_{\mu_i} \odot \mathbf{e}_{\mu_i}}{\boldsymbol{\pi}_{q|\mu}} \end{aligned}$$
The Ambiguity Reduction Metric ($\Delta \mathcal{F}$) is the difference between the baseline free energy and the memory-conditioned free energy:
$$\Delta \mathcal{F}(\mu_i) = \mathcal{F}(q(s_t), o(t)) - \mathcal{F}(q(s_t \mid \mu_i), o(t))$$
- If $\Delta \mathcal{F}(\mu_i) > 0$: The memory reduces cognitive uncertainty and actively resolves situational ambiguity.
- If $\Delta \mathcal{F}(\mu_i) \le 0$: The memory introduces cognitive conflict, cognitive dissonance, or redundant noise.
flowchart TD
Current["Current Mental State: q(s_t)"] --> Fusion["Bayesian Evidence Fusion with Memory e(mu_i)"]
subgraph UpdateMath ["Precision-Weighted Update Math"]
P1["Updated Precision: pi_new = pi_q + pi_mu"]
P2["Updated Mean: mu_new = (pi_q * mu_q + pi_mu * e_mu) / pi_new"]
P1 --> P2
end
Fusion --> UpdateMath
UpdateMath --> Cond["Conditional State: q(s_t | mu_i)"]
Cond --> Eval["Compute Conditional Free Energy F_conditional"]
Eval --> Diff["Compute Delta_F = F_baseline - F_conditional"]
Diff --> D_Pos["Delta_F > 0: Ambiguity Resolved (Elevated Score)"]
Diff --> D_Neg["Delta_F <= 0: Ambiguity Increased (Suppressed Score)"]
2.2 The Free-Energy Relevance Score (FERS)
To combine semantic relevance, information-theoretic uncertainty reduction, and homeostatic emotional alignment, Spector fuses these signals into the Free-Energy Relevance Score (FERS):
$$\text{FERS}(\mu_i \mid s_t, o_t) = \alpha \cdot \text{sim}(\mathbf{e}_q, \mathbf{e}_{\mu_i}) + \beta \cdot \sigma\left(\kappa \cdot \Delta \mathcal{F}(\mu_i)\right) + \gamma \cdot \mathcal{A}(\mu_i, s_t)$$
where:
- $\text{sim}(\mathbf{e}_q, \mathbf{e}_{\mu_i})$ is the baseline semantic similarity.
- $\sigma(x) = \frac{1}{1 + e^{-x}}$ is the logistic sigmoid mapping energy deltas into $[0, 1]$.
- $\mathcal{A}(\mu_i, s_t)$ is the Affective Resonance score derived from the Homeostatic Affective Core (mood-congruent retrieval).
- $\alpha, \beta, \gamma$ are non-negative weights satisfying $\alpha + \beta + \gamma = 1.0$ (default: $\alpha=0.50, \beta=0.35, \gamma=0.15$).
flowchart TD
subgraph MultiFactorFusion ["FERS Tri-Factor Fusion Engine"]
M1["Candidate Memory mu_i"] --> S1["1. Semantic Similarity\nsim(e_q, e_mu) [Weight: 0.50]"]
M1 --> S2["2. Epistemic Value\nsigmoid( Delta_F ) [Weight: 0.35]"]
M1 --> S3["3. Affective Resonance\nA(mu_i, s_t) [Weight: 0.15]"]
S1 --> WeightedSum["Weighted Linear Synthesis\nalpha * S1 + beta * S2 + gamma * S3"]
S2 --> WeightedSum
S3 --> WeightedSum
WeightedSum --> FinalScore["FERS Output Score in [0.0, 1.0]"]
end
3. High-Performance SIMD Vector Execution
In production enterprise deployments, memory recall must execute within sub-millisecond latency budgets. Evaluating transcendental logarithms ($\ln$), division by precision vectors, and quadratic deviations across 768-dimensional vectors cannot tolerate scalar loops or garbage collection pauses.
Spector Memory implements the mathematical kernel directly on SIMD 512-bit hardware vector registers (zmm0–zmm7), processing 16 floating-point values per clock cycle with zero memory allocations on the heap.
flowchart TD
subgraph MemoryRegisters ["512-Bit SIMD Register Pipeline (AVX-512 / 16 Floats Per Lane)"]
direction TB
R_QM["zmm0: qMean (Current Posterior Mean)"]
R_QP["zmm1: qPrec (Current Precision)"]
R_PM["zmm2: pMean (Generative Prior Mean)"]
R_PP["zmm3: pPrec (Generative Prior Precision)"]
R_OBS["zmm4: obs (Observation Vector)"]
R_OP["zmm5: obsPrec (Sensory Precision)"]
R_QM & R_PM --> FMA1["Vector Sub & Square:\n(qMean - pMean)^2"]
R_PP & R_QP --> DIV1["Vector Precision Ratio:\npPrec / qPrec"]
FMA1 & DIV1 --> KL_ACC["Vector FMA Accumulator (zmm6):\nD_KL Partial Divergence"]
R_QM & R_OBS --> FMA2["Observation Deviation:\n(qMean - obs)^2"]
FMA2 & R_OP & R_QP --> LIK_ACC["Vector FMA Accumulator (zmm7):\nLog-Likelihood Accuracy"]
end
KL_ACC & LIK_ACC --> REDUCE["Horizontal Vector Reduction (Sum Across SIMD Lanes)"]
REDUCE --> TAIL["Scalar Masked Tail Cleanup (i >= loopBound)"]
### Hardware-Level SIMD Execution Characteristics
By vectorizing the diagonal Gaussian Kullback-Leibler divergence and expected log-likelihood calculations directly on AVX-512 and AVX2 hardware registers, Spector achieves key operational characteristics:
* **$O(D)$ Vectorized Complexity**: Processes full 768-dimensional and 1536-dimensional embedding vectors in continuous 512-bit lanes (16 single-precision floats per CPU cycle) with masked loop tail handling.
* **Zero Heap Allocations**: Executes mathematical reductions and precision-weighted updates without allocating intermediate objects or triggering JVM garbage collection pauses.
* **Fused Multiply-Add (FMA)**: Leverages single-instruction hardware FMA operations to compute deviations and precision ratios concurrently in hardware registers.
---
## 4. Synaptic Relay Wiring & Pathway Integration
In Spector Memory, cognitive operations are structured as modular **Synaptic Relays** organized within the **Cognitive Pathway Engine**.
The `FreeEnergyGuidedRelay` is integrated into the recall cognitive pathway immediately following the initial multi-tier store scan and prior to associative graph expansion:
```mermaid
flowchart TD
subgraph RecallPathway ["Recall Cognitive Pathway Execution Flow"]
T0["1. TRANSDUCTION (Query Vector)"]
T1["2. PROSPECTIVE (Reminders)"]
T2["3. GOVERNED_RELEASE_GATE (Security Permissions)"]
A1["4. HOMEOSTATIC_BIAS (Affective Resonance)"]
T3["5. VECTOR_SEARCH (4-Tier Store Scan)"]
A2["⭐ 6. FREE_ENERGY_GUIDED (Active Inference ΔF Scoring)"]
T4["7. SCORING (Salience & Recency)"]
T5["8. GRAPH_EXPANSION (Hebbian & Temporal Links)"]
A3["9. HOPFIELD_ASSOCIATIVE (Attractor Gestalt Discovery)"]
T6["10. EVIDENCE_FUSION & BM25 / RRF"]
A4["11. MANIFOLD_RERANK (Riemannian Metric Geodesics)"]
A5["12. CONSTRUCTIVE_SIMULATION (Narrative Self-Model)"]
A6["13. CONSCIOUSNESS_CONTINUITY (Gaussian IIT Φ_CC)"]
T7["14. SORT_TRUNCATE & MMR / COLBERT"]
A7["15. CONSCIOUS_ACCESS (Global Workspace Bottleneck ~7)"]
T8["16. CONSOLIDATION (Post-Recall Epistemic Learning)"]
end
T0 --> T1 --> T2 --> A1 --> T3 --> A2 --> T4 --> T5 --> A3 --> T6 --> A4 --> A5 --> A6 --> T7 --> A7 --> T8
style A2 fill:#004B56,stroke:#00E6FF,stroke-width:3px,color:#fff
Relay Execution Semantics
flowchart LR
subgraph ExecutionStep ["FreeEnergyGuidedRelay Execution Lifecycle"]
E1["Receive RecallSignal\nCandidates (mu_1..mu_N)"] --> E2["Fetch Active Posterior q(s_t)\nfrom MentalStateTracker"]
E2 --> E3["Batch Evaluate ΔF\nvia SIMD FreeEnergyKernel"]
E3 --> E4["Compute FERS Scores\nFusing Semantic, ΔF & Affect"]
E4 --> E5["Append Diagnostic Breakdown\n(free_energy_delta, baseline_f0)"]
E5 --> E6["Emit Updated Signal\nto AssociativeGraphRelay"]
end
5. Comparative Evaluation: Pure Vector Search vs. Free-Energy Guided Recall
To evaluate the architectural impact of Active Inference on agent memory, consider an ambiguous multi-turn dialog scenario where semantic keywords alone fail to distinguish the relevant solution.
Scenario: Diagnosing an Incident Under Conflicting Constraints
- Sensory Query: "Production service latency spiked after deploying the caching layer. What was the root cause in previous incidents?"
- Candidate Memory Pool:
- Memory A (High Keyword Match / Low Epistemic Value): "Added Redis caching layer to production catalog service to decrease latency." (High semantic match, but zero diagnostic value toward the incident cause).
- Memory B (Moderate Keyword Match / High Epistemic Value): "Identified thread starvation and connection pool exhaustion when cache misses triggered simultaneous database fallback." (Directly explains the failure mechanism, maximizing uncertainty reduction $\Delta \mathcal{F} > 0$).
- Memory C (Irrelevant Distractor): "Meeting notes discussing the Q3 cache roadmap." (Surface overlap, but contributes negative epistemic value).
flowchart TD
subgraph Matrix ["Epistemic Value vs Semantic Keyword Overlap"]
direction TB
subgraph TopLeft ["Low Keyword Overlap, High Epistemic Value"]
Q2["Unrelated Domain Facts (Filtered Out)"]
end
subgraph TopRight ["High Keyword Overlap, High Epistemic Value (OPTIMAL TARGET)"]
MB["⭐ Memory B: Root Cause Analysis\n(Positive ΔF > 0, Resolves Root Cause)"]
end
subgraph BottomLeft ["Low Keyword Overlap, Negative Epistemic Value"]
MC["Memory C: Roadmap Distractor\n(Negative ΔF <= 0, Irrelevant)"]
end
subgraph BottomRight ["High Keyword Overlap, Low Epistemic Value (SURFACE TRAP)"]
MA["Memory A: Added Caching Announcement\n(High Cosine Match, Zero Diagnostic Value)"]
end
end
style MB fill:#003B46,stroke:#00E6FF,stroke-width:2px,color:#fff
style MA fill:#4A154B,stroke:#E01E5A,stroke-width:2px,color:#fff
Retrieval Comparison Matrix
| Retrieval Engine | Top-1 Retrieved Memory | Epistemic Quality | Agent Reasoning Outcome |
|---|---|---|---|
| Standard Cosine Vector DB | Memory A (Surface keyword match) | Fails to diagnose issue | Hallucinates or repeats that caching was added |
| Hybrid BM25 + Vector (RRF) | Memory A (Keyword overlap) | Explains the action, not the cause | Suggests checking if Redis is enabled |
| Spector AISME (FEGR) | Memory B (Max $\Delta \mathcal{F}$) | Identifies underlying cause | Accurately diagnoses connection pool starvation |
6. Architecture & Configuration Flow
Active Inference in Spector Memory is designed with a multi-tiered configuration hierarchy that ensures 100% backward compatibility and zero overhead when unconfigured.
flowchart TD
subgraph ConfigLayers ["Multi-Tier Configuration & Gating Flow"]
direction TB
C1["System Properties & Configuration\n(spector.properties / SpectorPropertyConstants)"] --> C2["AismeConfig Immutable Record\n(enabled, hyperparams, relay toggles)"]
C2 --> C3["SpectorMemory Engine Bundle\n(Initializes MentalStateTracker & Generative Model)"]
C4["Per-Query RecallOptions\n(Runtime Enablement & Dynamic Overrides)"] --> C5["RecallGates Specification\n(Evaluates Gating Predicate at Microsecond Scale)"]
C3 & C5 --> C6{"Gate Satisfied?"}
C6 -- Yes --> C7["Execute FreeEnergyGuidedRelay\n(Compute ΔF & Modulate FERS)"]
C6 -- No --> C8["Bypass Relay (Zero-Overhead Transparent Passthrough)"]
end
Runtime Decision Flow for Memory Gating
flowchart TD
R_Start["Recall Request Received"] --> G1{"Is aisme.enabled == true?"}
G1 -- No --> Legacy["Execute Standard Recall Pipeline\n(Zero AISME Overhead)"]
G1 -- Yes --> G2{"Is free-energy.enabled == true?"}
G2 -- No --> Legacy
G2 -- Yes --> ExecFE["1. Fetch Posterior q(s_t)\n2. Batch Evaluate SIMD FreeEnergyKernel\n3. Rescale Candidate Scores via FERS"]
ExecFE --> GraphRelay["Pass Candidates to Associative Graph Relay"]
Legacy --> GraphRelay
7. Summary & Architectural Takeaways
By grounding memory retrieval in the mathematics of Active Inference and Variational Free Energy, Spector Memory bridges the gap between biological neurocognition and high-performance software engineering:
- Epistemic Retrieval: Memories are ranked not just by how closely they mimic prompt text, but by their mathematical ability to eliminate situational uncertainty ($\Delta \mathcal{F}$).
- Dynamic Generative Priors: The agent's identity, core purpose, and cognitive profile shape retrieval precision, preventing persona drift and hallucination.
- Hardware-Accelerated Vector Execution: Pure SIMD vector kernels deliver continuous Gaussian variational inference with zero heap allocations on the hot path.
- Resilient Synaptic Pathways: Full modularity ensures seamless graceful degradation and zero overhead when running legacy workloads.
8. Scientific References & Research Foundations
The Active Inference Self-Model Engine (AISME) and Free-Energy Guided Recall (FEGR) are built on peer-reviewed foundations spanning computational neuroscience, information physics, and associative memory dynamics:
The Free Energy Principle & Active Inference:
- Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. doi:10.1038/nrn2787
- Friston, K., FitzGerald, T., Rigoli, F., Schwartenbeck, P., & Pezzulo, G. (2017). Active inference: a process theory. Neural Computation, 29(1), 1–49. doi:10.1162/NECO_a_00912
- Parr, T., Pezzulo, G., & Friston, K. J. (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press. ISBN: 9780262045353.
Interoceptive Homeostasis & The Self-Model:
- Damasio, A. (2010). Self Comes to Mind: Constructing the Conscious Brain. Pantheon Books / Vintage.
- Seth, A. K., & Tsakiris, M. (2018). Being a beast machine: The somatic basis of active inference. Trends in Cognitive Sciences, 22(11), 969–981. doi:10.1016/j.tics.2018.08.008
Continuous Modern Hopfield Energy Landscapes:
- Ramsauer, H., Schäfl, B., Lehner, J., et al. (2021). Hopfield Networks is All You Need. International Conference on Learning Representations (ICLR 2021). arXiv:2008.02217
- Krotov, D., & Hopfield, J. J. (2016). Dense associative memory for pattern recognition. Advances in Neural Information Processing Systems (NeurIPS 2016), 29, 1172–1180.
Information Geometry & Neural Population Manifolds:
- Amari, S. I. (2016). Information Geometry and Its Applications. Applied Mathematical Sciences, Vol. 194. Springer. doi:10.1007/978-4-431-55978-8
- Chung, S., & Abbott, L. F. (2021). Neural population geometry: An approach for understanding biological and artificial neural networks. Current Opinion in Neurobiology, 70, 137–144. doi:10.1016/j.conb.2021.10.010
Hierarchical Predictive Coding:
- Rao, R. P., & Ballard, D. H. (1999). Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79–87. doi:10.1038/4580
- Bastos, A. M., Usrey, W. M., Adams, R. A., et al. (2012). Canonical microcircuits for predictive coding. Neuron, 76(4), 695–711. doi:10.1016/j.neuron.2012.10.038
Integrated Information Theory ($\Phi$) & Global Workspace Theory:
- Tononi, G., Boly, M., Massimini, M., & Koch, C. (2016). Integrated information theory: from consciousness to its physical substrate. Nature Reviews Neuroscience, 17(7), 450–461. doi:10.1038/nrn.2016.44
- Dehaene, S., Lau, H., & Kouider, S. (2017). What is consciousness, and could machines have it? Science, 358(6362), 486–492. doi:10.1126/science.aan8871
- Graziano, M. S., & Webb, T. W. (2015). The attention schema theory: a mechanistic account of subjective awareness. Frontiers in Psychology, 6, 500. doi:10.3389/fpsyg.2015.00500
Experience Spector Memory
Spector Memory is open-source under the Business Source License 1.1 (converting to Apache 2.0).
- 🌐 Website: spectrayan.com
- 🧠 Homo Digitalis: homodigitalis.dev
- 💻 GitHub: github.com/spectrayan/spector