1. Vendor and Product Reality

Elomia Health operates a consumer-facing mental wellness chatbot under the Elomia brand, positioned in the same category as independent products such as Wysa, Woebot Health, and Youper. The product presents itself as a round-the-clock emotional support companion informed by cognitive behavioral therapy (CBT), with structured self-reflection exercises, mood tracking, and free-form conversation. The user proposition is access: someone struggling late at night, or someone unable to afford a therapist, or someone in a region without local mental health infrastructure, can have a supportive conversation immediately.

The architecture is the now-conventional pattern for therapeutic chatbots: a foundation language model conditioned on therapeutic training data and safety guardrails, augmented with a retrieval system over the user's prior sessions, mood-tracking inputs, and structured exercise outputs. Each session begins with a context-construction step that retrieves recent interactions, the latest self-reported mood ratings, and any flagged content from prior sessions, then conditions generation on that retrieved context. Crisis-detection classifiers run alongside generation, triggering escalation to crisis resources when self-harm indicators appear. The product serves a broad consumer base, and user reviews report it as genuinely helpful for everyday emotional support.

Elomia's strengths within this scope are real. The CBT-informed conversational quality is high. The escalation paths are responsibly designed. The mood-tracking and journaling features give users a self-reflection surface that has independent therapeutic value regardless of the chatbot interaction. The product is a sound implementation of the conversational-mental-wellness category as currently understood.

2. The Architectural Gap

The structural gap appears in the texture of longitudinal care. Elomia's model of the user between sessions is a corpus of retrieved facts: utterances, mood ratings, exercise outputs, timestamps. The model of the user during a session is whatever the language model reconstructs from that retrieved corpus inside the prompt context. There is no persistent affective representation: no quantitative emotional state vector that exists between sessions, evolves over time according to specified dynamics, and conditions the agent's posture independently of whatever happened to be retrieved.

This matters because emotional trajectories have dynamics that factual retrieval cannot capture. A user working through a difficult breakup over six weeks generates an emotional arc with specific shape: initial shock, cycling between anger and sadness, gradual acceptance interrupted by setbacks, slow rebuilding of emotional baseline. A human therapist tracks this arc with a felt sense of where the patient is now, which is informed by but not reducible to what the patient last reported. Elomia enters each session with whatever the retrieval pulled in. If the retrieved subset under-represents the underlying trend, the agent's posture is mis-calibrated to the actual emotional trajectory.

Mood tracking does not close this gap. The user's self-reported mood is an input to a persistent affective model; it is not the model itself. A persistent affective model integrates self-report with conversational signal density, session-frequency patterns, response-latency shifts, lexical affect indicators, and the interaction between multiple emotional dimensions. Without the model as a first-class object, these signals can only enter the system through whatever the retrieval-and-prompt pipeline happens to surface, which is contingent rather than structural.

Elomia cannot patch this from inside its current architecture because the architecture is built around stateless generation conditioned on retrieved text. Adding more retrieval, longer context windows, or summarization passes does not produce a persistent affective state any more than longer EHR notes produce a clinical relationship. The state has to exist as a deterministic computational object with named fields, governed dynamics, and runtime read-write semantics that the conversational layer queries and updates rather than reconstructs.

3. What the Affective State Field Provides

United States Patent Application 19/647,395 discloses the affective state field as a deterministic, policy-bounded data structure that encodes valence-weighted feedback derived from prior execution outcomes and environmental observations. It is not a generation-conditioning artifact reconstructed from retrieved text; it is a first-class field of the cognition platform, instantiated as a scalar valence value, a multi-dimensional vector encoding distinct modulation axes, or a structured record comprising named modulation fields. The field carries governed temporal dynamics: an update function bounded by policy, a decay curve, semantic hysteresis, and entropy-governed transitions. The disclosure is technology-neutral as to storage substrate and update implementation, so a skilled implementer can realize it over any datastore that supports the read-write and decay semantics described.

The distinguishing property, and the one the retrieval-conditioned pattern does not provide, is that affect modulates cognition without ever creating authority. The spec is explicit that the execution substrate provides resources without retaining authority over the semantic agent's cognitive state, and that the affective state field modulates disposition rather than granting permissions. Elevated internal state biases how candidates are weighed and how conservatively the agent commits; it does not by itself authorize any action. In a therapeutic deployment this is the difference between an assistant whose posture shifts because its governed state shifted and one that can be argued into an intervention it should not offer.

The named control fields disclosed are structural modulation axes, each mapped to a specific aspect of deliberation, not arbitrary labels. The spec enumerates, among others, an uncertainty sensitivity field (weighting uncertain inputs more heavily and biasing toward deferral or escalation when elevated), an ambiguity tolerance field (maintaining multiple candidate interpretations in parallel versus forcing early disambiguation), a novelty appetite field (willingness to engage previously unobserved patterns), a persistence-under-partial-failure field (retry and reformulate versus abandon), and an escalation-under-time-pressure field. A therapeutic mapping selects and interprets these axes for the domain, but the primitive being disclosed is the governed field and its dynamics, not any single label set.

Two further disclosed mechanisms matter for longitudinal care. First, the decay curve and semantic hysteresis mean the state evolves through elapsed time between sessions rather than resetting to whatever a retrieval pass happened to surface, and hysteresis prevents the posture from oscillating on transient signals. Second, the emotional quarantine lifecycle, disclosed as normal operation, a volatility detector, a quarantine state, a restricted mode, and hysteretic recovery, gives the agent a governed way to enter a conservative posture when affective volatility spikes and to recover only under defined conditions, rather than acting on a spike as if it were a stable reading.

4. Composition Pathway

Composition with Elomia preserves the company's differentiated layer entirely. The CBT-informed conversational quality, the consumer experience, the exercise library, the mood-tracking UX, the crisis-detection classifiers, the safety guardrails, and the distribution relationship with users all stay at Elomia. What is added underneath is the affective state field as a governed substrate. Each user's field state lives in a store that the application reads at session start, updates during the session through the policy-bounded update function, and lets evolve between sessions through the disclosed decay curve.

The integration points are well-defined and enumerated here so a skilled implementer can reproduce the approach. Session-start context construction reads field state alongside the existing retrieved corpus and conditions generation on both. Observed signals such as utterance affect classification, response latency, session frequency, and exercise completion feed the update function as structured observations. The crisis-detection layer is augmented rather than replaced: existing classifiers continue to run on utterances, and the field state adjusts how conservatively the assistant's posture is set around them, within policy bounds and subject to the emotional quarantine lifecycle when volatility is detected. Importantly, this modulation shifts disposition and sensitivity; it does not authorize the assistant to take an action a governing policy would otherwise withhold, and existing escalation rules remain the authority for crisis handling. Embodiments range from a single scalar valence carried across sessions, through a fixed vector of named axes, to a structured record with per-field decay constants and coupling, so the same architecture scales from a minimal consumer build to a richer supervised deployment.

The user-visible result is an assistant that re-enters each conversation with continuity rather than a fresh reconstruction. A user who cancels sessions during a period of elevated internal state returns to a posture that reflects the elapsed-time decay of that state, not a log entry that happens to mention the cancellation. A user whose signals have been drifting despite steady self-reports finds an assistant whose governed state tracks the drift the self-reports understate. The difference shows up at the longitudinal seam, which is where the retrieval-conditioned pattern is thinnest.

5. Commercial and Licensing Implication

One fitting commercial arrangement is an embedded substrate license. Elomia embeds the affective state field into its consumer applications, and a governed, auditable trajectory then also supports deployments that require supervision. To be precise about status: Elomia is a consumer wellness product, not a cleared medical device, and adding governed affective state does not change that classification or confer clinical validation. What it does provide is the governed, decay-aware state record that supervised and enterprise deployments tend to require, and that a retrieval log of utterances does not amount to.

Pricing aligns with the substrate role: a per-active-user license on the field store and update engine, with a separate enterprise tier where the governed trajectory carries supervision and audit obligations. What Elomia gains is a structural answer to the longitudinal-continuity gap that affects the retrieval-conditioned category broadly, expressed on an architectural axis rather than as a claim about any competitor doing its job poorly. What the user gains is continuity that survives the next model upgrade and the next retrieval-system change, and, where a human clinician is involved, a state record that person can actually review. Stated plainly, the affective state field does not replace Elomia's conversational layer; it gives that layer a persistent, policy-bounded emotional state to modulate against instead of reconstructing one each session. Affect modulates the assistant's disposition; it never becomes its authority.

6. Disclosure Scope

The technology attributed to the invention in this article, the affective state field as a deterministic, policy-bounded data structure with named control fields, a policy-bounded update function, a decay curve, semantic hysteresis, entropy-governed transitions, and an emotional quarantine lifecycle, together with the principle that affect modulates cognition without creating authority, is disclosed in United States Patent Application 19/647,395. This article is a dated public description of that disclosure and its application to affect-aware conversational assistants, and a skilled implementer may build the described approach over any storage substrate and update implementation consistent with the disclosed semantics.

All statements about Elomia, Elomia Health, and any other named product or company are external context describing the market as of the publication date. They are not claims of United States Patent Application 19/647,395, are drawn from publicly observable product behavior and the generally understood architecture of retrieval-conditioned therapeutic chatbots, and are offered only to locate the invention on an architectural axis. Product names are the marks of their respective owners and are used for identification only.