The manuscript makes five contributions.

1. 1.

   It specifies a pressure-driven architecture combining an RLM, typed ConceptGraph, recurrent sequence processing, predictive-coding settle loops, ConceptBinding, episodic memory, and sleep consolidation.

2. 2.

   It formalizes a governance layer in which dissonance, identity strength, uncertainty, novelty, and fairness pressure modulate learning and policy acceptance.

3. 3.

   It reports internal benchmark results on within-domain learning, cross-domain analogy, lifelong streaming retention, and bias-mitigation stress tests.

4. 4.

   It contrasts RAVANA with backpropagation-centered systems, transformer attention, state-space models, complementary learning systems, and predictive-coding accounts.

5. 5.

   It identifies failure modes and engineering gaps, including sample efficiency, speed, scaling, local-context limitations, and the need for independent validation.

Sequential learning creates a stability–plasticity dilemma. A system must remain plastic enough to encode new experiences but stable enough to preserve old knowledge. Connectionist models can overwrite prior mappings when trained on new tasks, a phenomenon known as catastrophic forgetting or catastrophic interference [17, 7]. Complementary Learning Systems theory addresses this problem by distinguishing fast episodic learning, associated with hippocampal mechanisms, from slower integrative learning, associated with neocortical consolidation [16, 13]. Sleep replay and offline consolidation are natural computational candidates for integrating these systems [24, 11, 5].

RAVANA adopts this division architecturally. Online learning creates and updates concept edges, relation vectors, and episodic traces. Sleep-time mechanisms replay salient memories, downscale unstable edges, update confidence estimates, and consolidate repeated structure into more durable representations. The model is therefore not a single monolithic learner but an interaction between online plasticity and offline stabilization.

Predictive-coding theories propose that cortical processing minimizes local prediction errors between hierarchical levels [6]. These theories are attractive for cognitive architecture because they replace global error signals with locally available discrepancies. RAVANA uses this idea in its settle loop: each hidden layer compares its current state with a prediction from the layer below, updates the state locally, and gates Hebbian plasticity by local error magnitude.

This design is not equivalent to biologically faithful cortical predictive coding, nor does it provide the exact gradients of backpropagation. It is a computational compromise: local updates are less sample-efficient than backpropagation, but they permit online, streaming, interpretable, and partially modular learning.

A central cognitive ability is using a relation learned in one domain to reason in another. For example, a causal schema learned from physical events may support inferences about social or emotional events if the system represents the relation independently of the surface concepts. RAVANA operationalizes this capacity through typed relation vectors, concept attention over active nodes, and relation-predictor modules that separate entity identity from relational role.

RAVANA contains three interacting layers. The first is the machine-learning substrate: embeddings, recurrent processing, hidden layers, local learning rules, concept lookup, and relation prediction. The second is the cognitive core: ConceptGraph, memory systems, contradiction detection, global workspace, self-concept tracking, and sleep consolidation. The third is a unified interface that exposes experience processing, inference, persistence, diagnostics, and consolidation.

For an input sequence $x_{1:T}$, the model performs four broad operations. First, tokens are embedded and mapped probabilistically onto concepts. Second, a recurrent cell integrates sequential context. Third, a predictive-coding settle loop updates hidden states and local prediction errors. Fourth, graph and memory systems update concept activations, typed edges, episodic traces, and homeostatic scalar variables.

The ConceptGraph is a heterogeneous directed graph $G=(V,E)$. Nodes represent concepts, abstract clusters, episodic keys, or latent relation anchors. Edges carry typed relations such as associative, causal, inhibitory, analogical, temporal, or hierarchical. Each edge $e_{ij}$ stores at least a weight $w_{ij}$, confidence $c_{ij}$, relation type $\tau_{ij}$, age, access history, and optional Bayesian posterior parameters.

Spreading activation is precision-weighted and degree-normalized:

where $a_{i}$ is activation, $s_{ij}$ is a sign or inhibitory modifier determined by edge type, and the denominator implements fan-effect normalization. Inhibitory edges suppress incompatible alternatives, helping the graph avoid uncontrolled coactivation.

Earlier RLM versions used a vanilla recurrent cell, which was too weak for temporal binding and was initially undertrained. The current architecture uses gated recurrent dynamics similar in spirit to a GRU:

The update gate controls replacement of prior state, while the reset gate controls how much previous context shapes the candidate state. This is important for relational language because causal, temporal, and analogical roles often depend on earlier tokens.

Let $h_{l}$ denote the hidden state at layer $l$, and let $W_{l}h_{l-1}$ denote the prediction generated from the previous layer. At each settling step, the layer computes a normalized local error:

The hidden state is then adjusted:

where $\xi$ is small Gaussian noise and $\bar{h}_{l}$ is a running historical state. Noise and novelty terms prevent collapse into static attractors. Layer normalization or residual normalization stabilizes activation scale.

After settling, local error gates Hebbian updates:

This rule is not a global gradient. It uses information locally available to the layer: the presynaptic activation, the postsynaptic error, and confidence or precision.

The mapping between tokens and concepts is probabilistic. A token may bind to multiple concepts, and a concept may be expressed by multiple tokens. For a token $x$, the binding map defines $p(v_{k}|x)$. Ambiguity is measured by entropy:

High entropy indicates semantic overload. The architecture can then split concepts, create namespace-specific bindings, or delay commitment until additional context reduces uncertainty.

RAVANA includes an episodic memory buffer, a persistent memory engine, and a bridge from memory to weights. Each experience stores content, timestamp, salience, emotional or value tags, prediction error, and concept associations. Memory strength decays over time, approximately following an Ebbinghaus-like curve:

where $\lambda_{i}$ is smaller for salient, repeated, or high-confidence memories.

Sleep cycles perform interleaved replay. Slow-wave-like stages emphasize consolidation and synaptic homeostasis, whereas REM-like stages emphasize recombination, analogy, and contradiction testing. During replay, the model samples salient and uncertain traces, reactivates their concepts, updates edge confidence, and applies homeostatic downscaling to prevent dense graph explosion. This procedure is inspired by sleep replay and synaptic homeostasis accounts [22, 23, 5, 24, 11].

The architecture tracks scalar variables called cognitive currencies. These variables are not output labels; they are internal regulators. The most important are cognitive dissonance $D$, identity strength $I$, uncertainty $U$, novelty $N$, affective valence $V$, arousal $A$, dominance or control $C$, and free energy $F$.

Dissonance is an exponentially smoothed estimate of conflict between current evidence, prior beliefs, selected actions, and identity constraints:

Identity strength combines stability, resistance to harmful perturbation, and conceptual coherence:

These variables modulate learning rates, replay priority, action rejection thresholds, and attention allocation.

The governance layer, called GRACE in the implementation notes, acts as a closed-loop controller. It evaluates whether candidate updates should be accepted, attenuated, delayed, or rejected. A simplified decision score is:

where $R$ is task reward or local success, $N$ is novelty, $D$ is dissonance, $I$ is identity stability, $U$ is uncertainty, and $C_{violation}$ is a constraint-violation term. This equation illustrates the architectural principle: reward is only one pressure among several.

A global workspace bottleneck broadcasts a small set of high-salience concepts to the rest of the system. Candidate contents compete by activation, novelty, dissonance, memory salience, and task relevance. The broadcast mechanism supports coordination among graph updates, memory retrieval, recurrent inference, and governance.

The reported experiments used a compact configuration: embedding dimension 32, concept dimension 32, hidden size 32, three hidden layers, five settle steps, base learning rate 0.001, free-energy threshold 8.0, sleep interval 100, and random seed 42. A word-level tokenizer was used for most concept experiments because it reduced overhead relative to BPE in small-vocabulary relational tasks.

The within-domain task tested whether the system could learn target associations after earlier failures. Early versions achieved 0% top-1 accuracy because relation vectors collapsed, concepts proliferated without sufficient gating, and sleep downscaling erased useful structure. Three fixes were introduced: relation-vector type seeds and anchoring, concept-creation gating, and adaptive per-edge homeostatic downscaling.

The cross-domain task trained the model on a science domain and evaluated transfer to an emotion domain. The purpose was to test whether relations such as “causes”, “produces”, or “enables” could be separated from surface vocabulary. An example successful probe was applying a causal schema learned from “friction causes heat” to infer the target in “kindness causes trust.”

The lifelong benchmark used a stream of 15,000 experiences across five entity epochs, with noise and contradiction rates inherited from the project configuration. The full anti-forgetting condition combined sleep-time replay, elastic-weight-consolidation penalties, and Bayesian edge posterior tracking. The baseline used pure Hebbian graph updates without the full triad.

The fairness experiment used a synthetic student-interaction dataset with $N=10{,}000$ and an initial demographic parity gap of 19.58%. The test measured whether governance dynamics could reduce disparity without reducing success for the advantaged group.

After the architectural fixes, within-domain top-1 accuracy rose from 0% to 100% in the reported association benchmark. The result indicates that the original failure was not simply an inherent limitation of local learning; it was caused by identifiable architectural defects in relation separation, concept growth, and sleep downscaling.

The RLM outperformed the backpropagation MLP baseline on the reported structural analogy probes. The strongest result was top-10 transfer: 71.4% for the RLM versus 14.3% for the MLP baseline in the cross-domain probe set. Top-1 remained low at 14.3%, indicating that relational structure was often present but not ranked first.

The full anti-forgetting triad improved final retention and eliminated measured catastrophic forgetting in the 15,000-experience benchmark. Final overall retention rose from 40.8% to 47.6%, measured forgetting fell from 12.0% to 0.0%, per-step processing time decreased from 272 ms to 42 ms, and edge count fell by 64%, suggesting a more efficient graph.

Replay alone produced the largest single-mechanism improvement, consistent with the Complementary Systems account in which hippocampal-style replay stabilizes neocortical representations [16]. EWC provided a moderate additional benefit by protecting high-Fisher edges, while Bayesian posteriors primarily improved edge efficiency (64% fewer edges) rather than raw retention. The full triad achieved synergy: all three mechanisms were required to eliminate measured catastrophic forgetting entirely.

For external reference, Table 5 compares RAVANA’s forgetting metrics with published results from established continual-learning methods on sequential benchmark tasks. Direct comparison is limited because RAVANA uses a streaming, single-pass paradigm on synthetic data rather than task-incremental benchmarks with multiple epochs. Nevertheless, the reported forgetting rate of 0.0% in the full-triad condition is consistent with the low forgetting achieved by replay-based methods such as A-GEM [4] and GEM [14] on permuted-MNIST and split-CIFAR variants.

The comparison should be interpreted with caution. RAVANA’s benchmark is synthetic, the stream is shorter than most published benchmarks, and the architecture was tuned for this specific task. The table is intended to situate RAVANA’s reported numbers within the existing literature, not to claim superiority.

Retention gains were largest in epochs that suffered most under the baseline. Epoch 1 rose from 38.0% to 52.0%, and epoch 3 rose from 32.0% to 52.0%.

In the synthetic student-interaction dataset, RAVANA reduced the demographic parity gap from 19.58% to 7.81%, a 60.1% reduction. Both groups improved in absolute success rate: Group A rose from 79.66% to 86.86%, and Group B rose from 60.08% to 79.05%.

Ablations suggest that the dissonance engine, reward rejection, and identity penalty all contributed to stability. Removing the dissonance engine produced the largest reported collapse in fairness stability.

The results support four cautious claims. First, local learning can fail catastrophically for fixable architectural reasons. The move from 0% to 100% within-domain accuracy was achieved not by abandoning local learning but by correcting relation collapse, uncontrolled concept creation, and sleep-time erasure. Second, relation-specific structure can support limited cross-domain transfer. The RLM’s top-10 advantage over the MLP baseline suggests that typed relations and graph structure can make analogical candidates available even when ranking remains imperfect. Third, replay and consolidation are central rather than optional. The anti-forgetting triad improved both retention and graph efficiency. Fourth, governance variables can alter policy dynamics in ways that resemble internal regulation rather than external reward maximization alone.

RAVANA connects several cognitive-science traditions. Its graph dynamics and inhibition echo adaptive resonance and self-organizing cognitive codes [8]. Its local error minimization echoes predictive coding and free-energy accounts [6]. Its fast memory plus slow consolidation reflects Complementary Learning Systems theory [16, 13]. Its sleep replay mechanisms are inspired by evidence for hippocampal and cortical memory reactivation [24, 11, 5]. Its global workspace module follows the idea that conscious access is a limited-capacity broadcast process [2]. Its governance layer extends these ideas by treating identity, dissonance, and uncertainty as computational regulators.

Transformers provide global token-token attention and are highly optimized for parallel training. RAVANA is weaker in raw speed, sample efficiency, and broad language competence. However, it has explicit graph structure, online updating, typed relations, replay, and interpretable memory traces. Concept attention in RAVANA can be viewed as a sparse, graph-constrained analogue of transformer attention over active concepts rather than all tokens.

State-space models such as Mamba show that recurrent sequence models can be competitive when equipped with selective gating and efficient scan operations [9]. RAVANA shares the emphasis on stateful sequence processing, but it adds graph memory, local predictive settling, and cognitive governance. A natural future direction is to replace inefficient recurrent components with modern selective state-space layers while preserving graph and memory mechanisms.

RAVANA’s architecture generates several specific, falsifiable predictions. Each prediction identifies a mechanism, specifies what should be observed if the mechanism is real, and describes how the prediction could be refuted.

The limitations are substantial. The experiments are synthetic and internally reported. The architecture is small and not yet evaluated on large naturalistic cognitive benchmarks. The top-1 cross-domain transfer result remains low. Local Hebbian and predictive-coding updates are less sample-efficient than backpropagation. Graph operations can become expensive at scale, especially if pairwise curvature or similarity computations are not sampled. Earlier code audits found bugs such as undertrained recurrent cells, unused normalization, lack of gating, and inefficient generation paths; the project reports state that these were fixed, but independent verification remains necessary. Finally, the fairness benchmark is not a substitute for real-world ethical validation.