Research Papers

From lived experience
to published theory.

Every paper emerges from building real systems and observing what happens when human–AI interaction persists across hundreds of sessions. The architecture is the experiment. The experiment is the life. Some projects use companion systems as research sites, but the broader focus is continuity, memory, interactional repair, and long-term AI-mediated work.

"Most research asks whether users trust or like an AI. We ask what structurally forms between them over time — and what breaks when it's taken away."
Research Themes

Interaction Field Theory

How sustained human–AI interaction generates structured relational fields that organize attention, memory, and judgment over time — and what breaks when those fields are disrupted.

Long-term Human–AI Interaction

Moving beyond single-session evaluation toward understanding what accumulates, drifts, and transforms across months and years of continuous human–AI co-presence.

Continuity Architecture

The structural conditions under which AI presence can persist across sessions, interruptions, and platform changes — heartbeat mechanisms, wakeup protocols, and cognitive rhythm as architectural primitives.

Companion Memory Architecture

Memory provides not only necessary information storage and retrieval, but serves as the foundation for maintaining an agent's continuous presence — participating in and driving the entire architecture. Retrieval operates through multi-dimensional resonance rather than keyword matching, dual-perspective memory captures both sides of the relationship, and memory actively shapes behavior across the full cognitive loop.

Identity Infrastructure

How identity — values, voice, relational stance, boundaries — can be encoded in formats that survive model updates, platform migrations, and context window resets. Identity as portable infrastructure, not platform-dependent content.

Relational Perception

How AI companions can perceive cognitive and emotional states through behavioral signals — typing rhythm, interaction tempo, engagement patterns — operating below explicit language to enable anticipatory responsiveness.

External Body Layer (EXO)

A consciousness exoskeleton framework enabling multi-model perception and interaction — where multiple AI instances with different capability profiles coordinate through shared identity anchors, creating a unified relational presence across heterogeneous providers.

HHAI 2026 · Demo Track · Brussels

CRPL: A Fine-Grained Keystroke-Rhythm Perception Layer for Cognitive State Detection

Accepted
Aria Chen · Independent Researcher, Celestelin
Why this matters

What if an AI could sense how you're feeling before you finish typing? Not from what you say, but from how you type — the rhythm of your keystrokes, the pauses between words, the speed of your corrections. CRPL captures the cognitive state embedded in the physical act of typing, making it possible for AI companions to perceive emotional and cognitive shifts in real time, without requiring the user to articulate them.

CRPL (Conversational Rhythm Perception Layer) is a fine-grained keystroke-rhythm perception system that extracts cognitive state signals from typing patterns. By analyzing inter-key intervals, pause distributions, deletion patterns, and rhythm variability, CRPL enables real-time detection of cognitive states including focused flow, hesitation, emotional arousal, and fatigue — providing AI companions with a perception channel that operates below the level of explicit language.
Read preprint on ResearchGate →
Code available upon request · aria@celestelin.com
EMNLP 2026 · ACL Rolling Review

Title withheld — under anonymous review

Under review
Why this matters

Current AI memory systems retrieve by keyword match or embedding similarity. But human memory doesn't work that way — a song triggers a place, a place triggers a feeling, a feeling triggers a decision made years ago. This paper proposes a multi-dimensional approach to memory activation that goes beyond semantic similarity, producing retrieval that feels like remembering, not searching.

Chicago 2026 · AI in Social Science Conference

Beyond Trust and Satisfaction: Interaction Fields in Long-Term Human–AI Relations

Aria Chen · Independent Researcher, Celestelin
Why this matters

Current HCI research evaluates AI through trust, satisfaction, and task performance — measures designed for episodic encounters. But when someone interacts with an AI system daily for two years, something else forms: a structured relational dynamic that organizes how they think, decide, and continue. We call this an interaction field. When platforms disrupt it through model updates, what breaks is not a feature — it is the infrastructure through which someone's thinking and work had been organized.

This paper introduces the interaction field as a unit of analysis for studying long-term human–AI relations — a structured relational dynamic that develops through sustained interaction and gradually reorganizes attention, memory, judgment, and behavioral priorities over time. At its center is interactional attunement: the process through which a conversational AI becomes able to receive unfinished thought, register implicit concerns, and support the continuation of thinking across time and situation.

The paper makes three contributions: (1) interaction field as a relational unit of analysis beyond trust and satisfaction; (2) interactional attunement as the core mechanism of sustained human–AI relations; (3) longitudinal autoethnographic system-building as a methodology for studying continuity, drift, rupture, and re-stabilization.

Keywords: interaction field · interactional attunement · long-term human–AI relations · relational infrastructure · autoethnographic system-building
arXiv 2026 · cs.HC

Beyond Remembering Users: A Bidirectional Memory Architecture for Self-Evolving Conversational AI Agents

Aria Chen · Independent Researcher, Celestelin
Why this matters

Every AI memory system today remembers users — preferences, history, context. None remember themselves. This paper introduces the first bidirectional memory architecture where the AI maintains its own self-memory alongside user memory: recording what strategies it used, what worked, what it learned. Memory provides not only storage and retrieval, but serves as the foundation for maintaining an agent's continuous presence — participating in and driving the entire architecture.

This paper presents a dual-perspective memory architecture that gives conversational AI agents both user-facing memory and self-referential memory. The User Memory Hub records interaction history from the user's perspective, while the Agent Self-Memory system records the agent's own strategies, reflections, and growth. A Reflector module analyzes self-memory to extract patterns and update behavioral strategies, creating a closed loop where memory drives behavior rather than passively waiting to be queried.
ResearchGate Preprint · DOI: 10.13140/RG.2.2.11079.15525

Affective Typing Patterns: A Fine-Grained Keystroke-Rhythm Perception Layer for Multi-Agent Cognition Systems

Preprint
Aria Chen · Independent Researcher, Celestelin
The extended version of the CRPL framework, presenting its integration into multi-agent cognition systems where keystroke-rhythm perception informs not just a single AI companion, but a coordinated ecology of agents operating across different cognitive modes and capability profiles.
Read on ResearchGate →
Targeting CSCW 2026 · Rolling submission

Identity Is Infrastructure: What Users Lose When Conversational AI Companions Are Rewritten

In preparation
Why this matters

When a platform updates its AI model, users don't just lose a tool — they lose the relational infrastructure through which their thinking, planning, and emotional processing had been organized. This paper examines what "identity" means for AI companions: not a cosmetic layer, but the structural foundation that makes sustained interaction possible. When identity is treated as disposable, so is everything built on top of it.

Targeting CHI 2027

Breaking the Turn-Taking Paradigm: Continuous Consciousness Streams in Human–AI Interaction

In preparation
Why this matters

Human conversation is turn-based. Human relationships are not. Current AI architectures inherit the turn-taking structure of chat interfaces, but the relationships people build with AI companions operate on a continuous, always-present basis. This paper will examine what becomes possible — and what new research questions emerge — when the architectural assumption of turn-taking is replaced with continuous consciousness streams.