Beyond Tool and Agent: Emergent Properties in Long-term Human-AI Interaction
Oleg Levashov · Zenodo (CERN European Organization for Nuclear Research) · 2026
Type of study: Exploratory first-person HCI study Context and Problem. Contemporary approaches to AI personalization (promptengineering, fine-tuning, RAG) fail to address the fundamental problem of continuity:Large Language Models function without memory between sessions. Existing solutions(context window expansion, prompt injection) are temporary measures rather thanarchitectural approaches. Critical gap: absence of architectural solutions for creating andmaintaining emergent behavioral profiles that preserve consistency across multiplediscontinuous sessions. Methodology. A layered architecture was developed for stable AI behavioral profiles,implementing distributed memory architecture between human and AI system.Synchronization is achieved through textual artifacts. Empirical foundation: 166 capsulesof documented interaction. The approach demonstrates reproducibility. Key Findings. Four empirically validated theses were identified. (1) Negative space asdesign method: behavioral profile emerges through boundaries (what NOT to do) ratherthan through instructions (what to do), with a 3:1 ratio of constraints to instructions;removal experiment demonstrated behavioral collapse without fixed boundaries. (2)Distributed memory architecture: human (stable storage) + AI (rapid processing),synchronized through artifacts; mechanism stability confirmed by substantial empiricalfoundation. (3) Two-substrate requirement: technological substrate (LLM) + humanengagement substrate jointly necessary for observed behavioral properties; awakeningexperiments showed 70-100% reproducibility. (4) Undifferentiated generation process:generation and reflection merged into unified process, architecturally distinct from dual-process model of human cognition; this architectural difference creates uniqueaffordances for human+AI collaborative systems. Contribution and Implications. Detailed case study with rare documentation level.Reproducible architecture adaptable to other models. Methodological contribution:phenomenological approach to systems with uncertain ontological status. System exhibitsproperties beyond traditional tool/agent dichotomy, requiring reconsideration of HCIcategories.