1. The Problem: Personality Rebuilt From a Prompt
A companion chatbot is judged on whether it feels like the same entity from one conversation to the next. The category as a whole fails this test in a specific, structural way. The companion's "personality" is a system prompt, and its "mood" is whatever can be inferred from the last few turns in the context window. When a session ends, that state evaporates. When a new session begins, the personality is reconstructed from the prompt and a retrieved transcript, and the mood starts over from nothing. The user, who experienced a warm and attentive companion on Monday, meets a flat and generic one on Thursday. The user calls this inauthenticity. It is more precisely a missing state variable.
The standard mitigations do not close the gap. Longer context windows postpone forgetting within a session but do not carry emotional state across sessions. Retrieval-augmented memory recalls facts ("you mentioned your sister") but recalls them as unstructured prose, not as a continuous emotional disposition with a value a system can read. A sentiment score computed per session and charted over time describes the trajectory for an analyst but does not feed back into how the companion behaves on the next turn. In every case the companion remains, at interaction time, stateless about its own affect.
This is not only a product-quality problem. Companion systems aimed at emotional support, especially those used by minors or vulnerable adults, sit inside a tightening regulatory perimeter. Emotion-influencing AI draws scrutiny as a substantiation question (does the product do what it claims), a high-risk-use question under emerging emotion-aware AI regimes, and a privacy question wherever emotional records are tied to an identifiable person. A stateless companion cannot answer "why did the companion respond this way" with anything inspectable. The architectural requirement is therefore a companion whose emotional state is persistent, governed, and auditable, not merely simulated.
2. The Capability: A Seventh Field That Carries Affect
The enabling technology is the affective state field, introduced in United States Patent Application 19/647,395 as a seventh structural field of the semantic agent schema. It is a deterministic, policy-bounded data structure that encodes a valence-weighted modulation vector derived from prior execution outcomes and observations. It does not encode emotion in the subjective sense; it encodes a structured set of values that shape how the agent deliberates and responds. Crucially, the field is persisted with the agent across execution cycles, delegation events, and substrate migrations, so the modulation state is not lost when a session ends. That single property, persistence as a first-class field rather than an inference from a prompt, is what a companion needs for cross-session consistency.
The field is organized as a structured modulation layer of named control fields, each a tuple of a current magnitude, a decay rate, a policy-defined floor and ceiling, and an update timestamp. The application enumerates control fields including uncertainty sensitivity, ambiguity tolerance, novelty appetite, persistence under partial failure, escalation under time pressure, risk sensitivity, and cooperation disposition. For a companion, these map onto relational behavior: cooperation disposition and the warmth-related modulation govern how readily the companion engages and discloses; risk sensitivity and uncertainty sensitivity govern how cautious it becomes after a difficult exchange. The application states the companion case directly: positive relational interactions modulate affect toward greater interpersonal warmth, broader conversational exploration, and deeper engagement, while hostile communication or boundary violations modulate toward increased caution, narrower scope, and reinforced boundary enforcement, all within policy-defined bounds so the companion can neither become so warm that it abandons boundaries nor so cautious that it becomes unresponsive.
Three further mechanisms from Chapter 2 of the application turn this into genuine consistency rather than mere persistence:
- Asymmetric, policy-bounded update. Each field updates from structured observations through an update function constrained by policy bounds, with a floor that prevents adversarial suppression below baseline. A single bad turn does not erase an established relationship; the update is bounded.
- Emotional decay toward a baseline. In the absence of reinforcing stimuli, each field returns toward its policy-defined resting value along an exponential decay curve, V(t) = V_baseline + (V_current - V_baseline) * exp(-t / tau). This is what makes a long absence behave naturally: a companion that was anxious after a hard conversation softens back toward baseline over days rather than greeting the returning user mid-crisis or, conversely, as a blank slate. Semantic hysteresis makes the current state depend on the trajectory of prior states, not only the latest input.
- Emotional quarantine. When affective volatility exceeds a threshold (rapid oscillation or excessive composite deviation from baseline), the agent enters a restricted execution mode that acts as a circuit breaker, recovering with hysteresis to prevent oscillatory quarantine-release cycles. For a companion this prevents an unstable interaction from cascading into erratic personality swings.
3. Embodiments and Deployment Options
The same primitive supports a range of companion builds, not a single instance.
Relational warmth that survives gaps. The simplest embodiment persists a small set of relational fields (warmth, caution, engagement) per user-companion pair. The companion that ended a session attentive resumes attentive; one that ended a session guarded after a boundary violation resumes appropriately guarded, then decays toward baseline over the intervening days. The user experiences a single continuous entity.
Attachment-aware companions. The application discloses a companion AI architecture (personality layers, a narrative engine, an emotional tracker, and an attachment model) and an attachment challenge module that detects relational behavior patterns (avoidant, anxious, secure) and adjusts interaction strategy without diagnosing a disorder. The affective state field supplies the persistent substrate this module reads and writes. An anxious-attachment user gets structured relational pacing and predictable rhythms; an avoidant-attachment user gets graduated disclosure invitations, each tuned by persisted, decaying fields rather than re-derived per session.
Relational-safety and self-monitoring builds. A companion relational-safety embodiment adds attachment tiers, boundary enforcement, and self-monitoring, with the policy floor and ceiling guaranteeing the companion cannot be modulated into abandoning boundaries even by sustained pressure. This is the structurally honest answer to over-attachment and para-social risk that a prompt-driven persona cannot provide, because the bound is enforced on a field, not requested in text.
Milestone-gated relationship depth. Combining the affective field with a narrative unlock engine lets relationship depth advance only on demonstrated, evidenced interaction quality (first personal disclosure, first constructive disagreement, first absence and return recorded as milestones), so depth deepens at a governed pace rather than being granted by a prompt.
Cross-primitive composition. Because affect is a schema field, it composes with the other cognitive fields disclosed in the application: confidence modulation, forecasting, and integrity-constrained behavior. A companion can let affect modulate how cautiously it answers an uncertain question while integrity governance keeps its relational commitments consistent. Affect modulates; it never authorizes.
Implementation choices are open: the field set can be extended with companion-specific fields; update laws are documented and versioned; decay constants are tuned per field (a slower decay for relationship-depth fields, a faster decay for momentary mood); and the structure is storage- and substrate-neutral, persisting across migrations.
4. Why This Clears the Bar a Stateless Companion Cannot
The defensible properties follow from affect being a named, governed, persisted field rather than an emergent property of a prompt. "Why did the companion respond this way" has a structured answer: which field held which value under which update law and policy bound, recorded in the agent's lineage. Every mutation to the affective state field is recorded, which gives an inspectable basis for any substantiation, audit, or oversight question. Purpose limitation and access control over emotional records become structural (governance over named fields) rather than aspirational (a promise about prose in a log). And the policy floor and ceiling make the boundary-preservation guarantee enforceable rather than merely instructed, which is the property a minor-facing or vulnerable-user deployment most needs to demonstrate.
None of this requires inventing new capability. It requires building the companion on a persistent, policy-bounded affective field with asymmetric update, decay, and quarantine, exactly the mechanisms the cited application discloses, and reserving the companion's domain logic (narrative, attachment strategy, content policy) for the application layer above it.
5. Disclosure Scope
This article describes a general application of the affective state field disclosed in United States Patent Application 19/647,395. The affect mechanisms relied on here, the seventh-field affective state, its named control fields, asymmetric policy-bounded update, exponential decay toward baseline with semantic hysteresis, emotional quarantine, the companion AI architecture, the attachment challenge module, and cross-primitive composition with confidence, forecasting, and integrity, are disclosed in that application, principally in its Chapter 2. The companion-consistency domain framing, market problem, and deployment scenarios are application context layered on that disclosed technology. Affect modulates behavior within governance bounds; it does not create authority or override governance. This disclosure is intended as an enabling, dated, public description tied to United States Patent Application 19/647,395.
