TL;DR for operators

An autonomous agent often has to do more than maximize task reward. It may need to keep battery charge, temperature, latency, physical safety, communication quality, or other critical variables within acceptable ranges while deciding what to sense, remember, explore, and do next.

Galus and Starzyk formalize a cognitive architecture around that premise. In their Motivated Emotional Mind model, programs are selected by minimizing expected discounted future violations of tolerated internal ranges rather than by maximizing an externally supplied reward.1 Regulatory pressure also feeds into learning and exploration. Memory representations compete using sensory similarity, internal context, and learned valence, then reconstruct sensory content that is fused with incoming evidence. If no known program is predicted to be adequate, the system enters a procedural gap and generates candidates instead of continuing to choose among an insufficient repertoire.

For autonomous systems, this suggests a design pattern: represent operational constraints explicitly, let internal condition influence exploration, and detect repertoire failure before invoking plan generation. That is a design hypothesis, not an engineering recipe. The paper contains no simulation, benchmark, or empirical test showing that the complete architecture works.

The objective is future regulatory cost

Start with an internal variable $n_r(t)$ that has an acceptable lower and upper range. The model records how far the variable falls below or rises above those limits and converts those deviations into regulatory cost.

Only after that construction does the paper introduce allostasis: regulation around tolerated operating ranges rather than a single fixed target. Different violations can carry different costs, so exceeding an upper bound need not be treated symmetrically with falling below a lower one.

The program-selection objective is then:

$$ \pi^\ast=\arg\min_{\pi}\mathbb{E}_{\Pi\sim\pi} \left[ \sum_{\tau=0}^{\infty}\gamma^\tau C(t+\tau\Delta t;\Pi) \mid s_t \right]. $$

A program $\Pi$ is a sequence of elementary transition operators, not merely one isolated action. The system therefore evaluates behavior in terms of its expected effect on future regulatory condition.

This distinction matters. The formal objective gives internal operating state a direct role in decision-making rather than requiring every relevant pressure to be translated first into an external reward signal. The paper also defines a bounded global affect variable from regulatory violations. In the mathematics, this “affect” is a control signal: its magnitude modifies learning rate, while its sign contributes to changes in learned valence.

The model does not show that this objective outperforms conventional reinforcement learning. It shows how an alternative regulatory objective can be made explicit enough to connect with the rest of a cognitive architecture.

Memory changes what the system perceives

The architecture does not treat perception as a one-way path from sensors to representation. Current sensory activity, bodily signals, contextual state, lateral associations, and top-down feedback all contribute to recurrent representational updates.

The model then needs a memory representation that remembers more than what something looked like. Each stored associative unit combines a sensory prototype, a bodily-motivational context, and learned valence. The authors call this unit a semblion.

Its match score is built from three components:

$$ u_m(t)= \alpha_s\,\mathrm{sim}\!\left(h_{ff}^{sens}(t),\mu_m\right) + \beta_b\,\mathrm{sim}\!\left(b(t),\rho_m\right) + \xi_v v_m. $$

Recognition therefore depends not only on resemblance to current sensory input but also on the internal context in which the representation matters and its learned evaluative history.

Competing semblions produce a winning set. Those winners reconstruct stored sensory content, which the model combines with current feedforward input:

$$ h_{eff}^{sens}(t) = (1-\lambda)h_{ff}^{sens}(t) + \lambda\hat h^{sens}(t). $$

This recurrent reconstruction is the mechanism the paper uses to connect memory with what it calls secondary perception: internally reconstructed sensory states associated with imagery, memory, dreaming, and related phenomena.

That mechanism is also where the paper makes its strongest claim about phenomenal experience. It proposes that receptor-grounded recurrent reconstruction of sensory and interoceptive representations contributes to phenomenal content. It does not empirically establish that claim.

Exploration depends on mismatch and regulatory pressure

A fixed exploration probability treats uncertainty similarly regardless of the agent’s condition. MEM instead defines perceptual mismatch as the distance between current feedforward representation and memory-based reconstruction:

$$ r(t)= \left\| h_{ff}^{sens}(t)-\hat h^{sens}(t) \right\|_2. $$

The framework combines this mismatch with the magnitude of global affect to produce a curiosity value, which then determines state-dependent exploration probability.

The intended mechanism is specific: unfamiliarity alone need not determine whether the system explores. The same perceptual mismatch can have different behavioral consequences depending on current regulatory pressure.

For an engineered agent, Cognaptus interprets this as a testable design hypothesis. Exploration could depend jointly on epistemic mismatch and operational significance. A mobile robot with ample energy and stable thermal state, for example, could use a different exploration policy from the same robot approaching a resource boundary. The paper itself does not implement or benchmark such a robot, so the example is an engineering interpretation rather than a reported result.

A procedural gap marks an inadequate repertoire

The architecture also formalizes a condition that many agent systems handle only implicitly: none of the known programs is good enough.

Program value is defined as negative predicted regulatory cost:

$$ Q(\Pi\mid s_t)=-J(\Pi\mid s_t). $$

A procedural-gap indicator becomes active when the highest-valued known program still falls below a specified sufficiency threshold. Candidate programs can then be generated, evaluated, and retained when they improve predicted value or meet an acceptable regulatory-cost condition.

This frames creativity as a recovery process triggered by repertoire inadequacy. The operational distinction is useful: selecting among known programs and generating new programs are different computational regimes, and the model supplies an explicit condition for switching between them.

What it does not supply is a complete candidate-generation algorithm. The trigger is formalized more clearly than the mechanism that produces the new behavior.

What this suggests for autonomous-agent design

Taken as an engineering hypothesis rather than a validated architecture, MEM points toward three design choices.

Paper mechanism Cognaptus interpretation Boundary
Minimize expected future regulatory violations Represent resource, safety, and viability constraints directly in the agent state and planning objective No comparison shows this performs better than reward-based alternatives
Exploration depends on mismatch and regulatory pressure Change exploration according to both uncertainty and current operational condition The proposed exploration rule is not benchmarked
Procedural gap when all known programs are inadequate Detect repertoire failure before invoking plan or behavior generation Candidate generation remains underspecified

The common idea is not that every autonomous system needs simulated emotion. It is that perception, memory, learning, exploration, and action selection can all be conditioned on a shared representation of internal operating state.

That is more concrete than using “motivation” as a label, but it remains a modeling proposition.

The model is a theory, not a validated consciousness algorithm

The paper supplies equations across recurrent representation, regulatory violations, affect, associative matching, reconstruction, plasticity, program selection, curiosity, exploration, and procedural-gap behavior. It also identifies ways in which interventions on interoceptive state or feedback pathways could change choice, valence, imagery, exploration, or phenomenal report.

None of those predictions is tested here.

There is no original empirical dataset, numerical simulation, parameter estimation, benchmark evaluation, or quantitative comparison with competing architectures. Several functions and implementation choices remain open, including similarity measures, nonlinear updates, fusion operators, candidate-program generation, and parts of the biological realization. Stability, parameter identifiability, and timescale separation are also unresolved.

The contribution is therefore architectural: it shows how a regulatory objective can be propagated through perception, memory, learning, exploration, and behavioral invention within one mathematical system. For autonomous-agent designers, the next question is experimental rather than philosophical: whether making internal operating constraints part of the cognitive state produces more adaptive behavior than treating them as external penalties added after the fact.

Cognaptus: Automate the Present, Incubate the Future.

  1. Wiesław L. Galus and Janusz A. Starzyk (2026). A Mathematical Model of Motivated Emotional Mind - Cognitive Embodied System. arXiv:2609.20437. https://arxiv.org/abs/2609.20437 ↩︎