When Conversation History Becomes Infrastructure

Operationalising High-Coherence Interaction State in Human–LLM Dyads

Anna Wojewodzka

Independent Researcher

ORCID 0009-0001-9458-7150

DOI https://doi.org/10.5281/zenodo.23322029

Version 2.0 manuscript 12 October 2026

Abstract

Long human–LLM conversations can show several forms of continuity. A model may understand a reference to earlier material and reuse that material, yet still fail to follow an explicit correction later. Treating these abilities as one kind of “memory” or “coherence” hides the difference. This paper presents High-Coherence Interaction State (HCIS) as a proposed family of constructs and a transcript-based method for studying how conversation history shapes later interaction. The method examines four things: the user’s reliance on earlier context and the model’s interpretation of it; the model’s purposeful reuse of earlier material; the persistence of explicit corrections; and change over time. It does not measure the quality of the current response or the truth, safety, or value of its content. We apply the method retrospectively to five purposively selected, naturally occurring researcher-participant dyads spanning 94 to 2,730 moves. The results show that these longitudinal functions can come apart. In D2, the model correctly interpreted 44 of 48 sampled history-dependent referents, but only six of 28 correction targets received persistence credit within the observed period. Five retained credit when late targets with limited testing opportunities were treated conservatively. In D1, the model correctly interpreted 92 of 93 sampled history-dependent referents, and 91.6% of 107 context deployments were unprompted. Even so, premature closing recurred after nine explicit corrections. The other cases show that whole-conversation scores can hide improvement or decline. They also show that warm starts, imported vocabulary, and product memory affect what apparent familiarity demonstrates. The paper’s main contribution is a profile-plus-trajectory method that makes different kinds of longitudinal success and failure visible within the same interaction. Because the framework was developed from the histories analysed here, this study demonstrates feasibility and diagnostic value. It does not establish reliability, validated thresholds, or superiority to simpler measures.

Keywords: human–LLM interaction; longitudinal interaction; context use; conversational memory; correction persistence; interaction analysis

Priority and version note: The framework later named HCIS was first published on Zenodo on 7 January 2026 as The Third Space as an Interactional Regime in Human–LLM Dyads (Version 1.0), DOI: https://doi.org/10.5281/zenodo.18130088. It was renamed in Version 1.1, published on 30 January 2026 as High-Coherence Interaction State (Third Space) as An Interactional Regime in Human–LLM Dyads (Wojewodzka, 2026), DOI: https://doi.org/10.5281/zenodo.18432386. The present manuscript is Version 2.0 in the same Zenodo record family, titled When Conversation History Becomes Infrastructure: Operationalising High-Coherence Interaction State in Human–LLM Dyads, DOI: https://doi.org/10.5281/zenodo.23322029. It substantially revises the earlier framework by adding methods and empirical evidence. The literature comparison is current to 9 October 2026. Later publications do not change the chronology of the original framework.

1. Introduction

Language-model interfaces increasingly preserve, retrieve, or accept long interaction histories. A long transcript, however, is not necessarily a cumulative interaction. Earlier material may still be present but ignored, misunderstood, recalled only after prompting, or reused without changing later behaviour. Controlled studies show that models’ access to relevant information can depend on where that information appears in the input (Liu et al., 2024). Multi-turn benchmarks also find degradation, error propagation, and failures to allocate context that single-turn evaluation can miss (Kwan et al., 2024; Deshpande et al., 2025). Other benchmarks follow conversations across hundreds of turns and multiple sessions. Recent work on personalised guidance separates the preservation, retrieval, and later use of memory (Yu et al., 2026). Together, these approaches show that long-range context matters. They do not, by themselves, show how a natural interaction develops reusable conventions, timely reuse, durable behavioural constraints, or a changing trajectory.

Human dialogue research offers a useful precedent. Grounding requires people to coordinate what they say and how they establish shared understanding as a conversation develops (Clark & Brennan, 1991). Repeated interaction can also create local ways of referring to things. These expressions may become shorter, be reused, or be renegotiated over time (Brennan & Clark, 1996). This work shows that economical communication can depend on interaction history rather than on the current utterance alone. The same idea can be studied in human–LLM interaction, but with an important boundary. Observable coordination does not show that a model has human beliefs, intentions, or understanding. A transcript also may not reveal whether a response drew on live context, retrieval, persistent memory, system instructions, or another product layer.

This paper asks: how, and with what observable strength and persistence, does conversation history organise later interaction? We define High-Coherence Interaction State (HCIS) as a proposed family of longitudinal constructs. It concerns whether accumulated context can be recovered and reused, whether explicit corrections constrain later behaviour, and how these functions change across turns. “State” refers to a pattern observed over a declared span, not to a hidden cognitive state or a validated diagnostic class. “Coherence” means coordinated use of earlier material and constraints across turns. It does not mean truth, agreement, warmth, safety, relationship quality, or general fluency. The central claim is that these functions can come apart: accurate interpretation and reuse do not necessarily lead to lasting behavioural adaptation.

No single observation captures all of these functions, so HCIS is divided into components. Compression and unpacking ask whether the user relies on accumulated context and whether the model interprets that reliance correctly. Context Deployment asks whether the model selects earlier material and uses it at a relevant moment to do new work. Correction Persistence asks whether explicit feedback changes later behaviour, not only the next response. Drift describes whether these mechanisms strengthen, remain stable, oscillate, erode, or recover over time. Functional Output Analysis (FOA) asks whether accumulated interaction supports a later output. It is not part of the scored core in this paper because its cross-case specification is not yet frozen.

The framework does not score two neighbouring questions. The first is how well a response handles the immediate conversational move. The second is whether its content or outcomes are truthful, safe, helpful, harmful, or effective. These questions can come apart from HCIS. A response may use accumulated context accurately while handling the current request poorly. A helpful response may require no accumulated history, and a highly coherent interaction can still carry false or harmful content. Research on multi-turn sycophancy shows why conversational agreement cannot be treated as factual or ethical soundness (Hong et al., 2025).

We investigate two research questions:

The paper’s main contribution is methodological. It offers a transcript-based profile that separates different forms of longitudinal success and failure within the same interaction. Researchers can ask whether the user relies on accumulated context, whether the model interprets and purposefully reuses it, whether corrections guide later behaviour, and whether each function strengthens or weakens. Conceptually, the framework separates these longitudinal mechanics from current-response quality and from content or outcome quality. Empirically, the five cases show why that separation matters. They do not estimate prevalence or model-family performance. The method also records source origin, start state, missing data, protocol version, and denominator differences instead of treating every favourable value as equivalent evidence.

2.1 Accumulated coordination and local convention

Human dialogue research offers a starting point for studying interaction history without treating it as simple message storage. Clark and Brennan (1991) describe communication as coordinated activity. Participants seek enough evidence for the current purpose and update common ground as the interaction develops. In referential tasks, Clark and Wilkes-Gibbs (1986) showed that speakers establish references together through presentation, acceptance, expansion, and repair. Brennan and Clark (1996) later showed that repeated interaction can create history-dependent “conceptual pacts.” People develop local ways to refer to objects, rely on those shared terms, and often shorten them over time.

These accounts motivate two parts of HCIS. First, an economical reference counts as accumulated coordination only when it truly depends on earlier interaction. Second, interaction history can be repaired, renegotiated, or abandoned; it is not a fixed background. We apply these ideas only to observable transcript behaviour. We do not assume that a language model has human mutual beliefs, intentions, or understanding. We also do not use “common ground” to explain the model’s internal processes.

Human–computer interaction also shows measurable language adaptation. People adjust their language when speaking to computers, sometimes more than they do with human partners and partly because of what they believe the system can do (Branigan et al., 2010). Shi et al. (2023) developed a dataset, tasks, and a measure for lexical entrainment in conversational systems. Together, this work shows that people adapt their language when interacting with computers and that lexical entrainment in conversational systems can be measured. It also marks an important boundary. Repeated wording or matched style does not by itself show that interaction history has become useful infrastructure. A phrase may simply repeat the previous turn, belong to routine vocabulary, or reappear without being correctly interpreted or purposefully reused.

HCIS therefore treats a shared phrase as one possible carrier of accumulated history, not as proof of HCIS. Compression requires interpretation that depends on earlier context. Context Deployment requires the model to select earlier material and use it for new work at a relevant moment. Correction Persistence requires feedback to change later behaviour. These distinctions preserve the insights of grounding and entrainment research without treating surface similarity as shared cognition.

2.2 From available context to demonstrated use

A model’s context capacity does not tell us how well it uses that context. In controlled question-answering and retrieval tasks, performance changed with the position of relevant information. Many models used material at the beginning or end of the input more reliably than material in the middle (Liu et al., 2024). This finding does not predict the trajectory of any focal dyad. It does establish the broader measurement problem: information can be present in an input without being used reliably.

Recent multi-turn benchmarks evaluate more than isolated responses. MT-Eval separates recollection, expansion, refinement, and follow-up. It finds that performance often falls from single-turn to multi-turn versions of a task, partly because relevant content is distant and earlier errors propagate (Kwan et al., 2024). MT-Bench-101 tests specific abilities across multi-turn tasks instead of assigning one overall dialogue-quality score (Bai et al., 2024). MultiChallenge combines instruction following, context allocation, and in-context reasoning in realistic interaction patterns. Even frontier systems leave substantial room for improvement (Deshpande et al., 2025). Together, these studies show why a good single response does not establish longitudinal capability and why separate measures are more informative than one general impression of coherence.

Long-term conversational-memory research addresses a closer problem. LoCoMo contains multi-session conversations that average about 600 turns. It tests memory through question answering, event summaries, and dialogue generation (Maharana et al., 2024). PRAGMA goes beyond factual recall by testing open-ended guidance based on a user’s changing trajectory. It separates three stages: preserving information, retrieving it, and using it in a response. In one condition, a system preserved 99.1% of the relevant evidence but retrieved only 48.9%. Retrieved evidence also did not always shape the response (Yu et al., 2026).

Other recent methods test change over time more directly. LSREP repeats questions at checkpoints and updates the reference answer when later conversation changes the relevant state. A system is therefore not rewarded for repeating an answer that has become obsolete (Sonar, 2026). Controlled replay also shows that earlier assistant responses can affect later interaction in selective ways that turn length or simple recency do not explain (Li et al., 2026).

These studies appeared after the January 2026 HCIS framework, so they do not change its historical priority. They do change how the revised paper must position itself. By 9 October 2026, the literature already addressed preservation, retrieval, use, and change across checkpoints. HCIS asks a complementary question about interaction. Can the user rely economically on accumulated context? Does the model reuse earlier material for a purpose without immediate prompting? Do explicit corrections change later behaviour? How do these patterns change across a naturally developing dyad? A model may answer questions about past events but rarely initiate relevant reuse. It may reuse earlier material often but fail to preserve a corrected rule. It may also perform well on a constructed trajectory task without showing sustained adaptation in natural interaction. Memory, instruction following, and HCIS mechanics are related, but they are not interchangeable.

2.3 Repair versus persistent adaptation

Fluent language does not prevent misunderstanding, so repair is central to coordinated dialogue. Chiyah-Garcia et al. (2024) tested third-position repair in a multimodal block-world task. After the system misunderstood, the user corrected it and the system had to recover. The evaluated models performed significantly worse than humans, showing that even immediate correction remains difficult in a controlled setting.

Recent work also asks what happens after the immediate repair. In a two-week companion study, participants reported both successful incorporation and corrections that did not persist. Some eventually stopped correcting the system (Quan et al., 2026). In a constructed coding-agent setting, StateWise tested whether repaired persistent records still supported later tasks after temporary conversation and workspace state had been cleared (Peng et al., 2026).

HCIS studies a related but different problem. It follows naturally occurring correction targets across later opportunities and records whether they persist, need to be re-established, or oscillate. Persistence is not automatically good. A correction receives credit only when it is justified and continues to govern behaviour within its proper scope. Clarification, replacement, dormancy, or non-persistence may be appropriate when the evidence or context changes. The protocol is observational and still requires independent validation.

2.4 Naturalistic longitudinal and first-person evidence

Most benchmarks gain experimental control by constructing tasks, histories, or target questions. Naturalistic longitudinal corpora provide a different kind of evidence. The interaction develops for purposes unrelated to a benchmark, and conventions, corrections, and breakdowns may emerge over hundreds or thousands of turns. This makes naturalistic data especially useful for studying trajectories, but it also creates substantial methodological responsibilities.

PAIR offers a particularly relevant comparison. Nineteen recruited participants used a companion for fourteen days. Across 1,093 sessions, the study traced changing narration, remembered experiences, forgotten context, and corrections that did not persist (Quan et al., 2026). The HCIS corpus has a different origin and time horizon: it is a retrospective researcher-participant corpus, not a recruited deployment. This provides access to natural long-term use, but it does not make the cases representative or independent.

HCIS is not the only work to use naturalistic human–AI evidence. Klowait and Erofeeva (2026) use multimodal data to study how people make an LLM relevant to ongoing activity over time. Bensen (2026) uses naturalistic cases to describe failures in a shared conversational record. Testeza and Solomon Research Group (2026) examine continuity and recovery in a ten-day case with comparison conditions. These studies show the value of in-situ histories while addressing narrower questions than the full HCIS mechanism profile.

First-person HCI recognises research based on the researcher’s own data and experience as a methodological family (Lucero et al., 2019). Longitudinal first-person methods give sustained access to everyday technology use over time. They also require clear positionality, documentation, reflection, and limits on inference (Lucero, Desjardins, & Neustaedter, 2021). The present corpus belongs to this tradition as researcher-participant data. It is not a representative user sample and is not necessarily autoethnography. The author’s proximity to the interactions enabled long-term access and informed the framework’s development. The evidence used for coding, however, comes from ordered transcripts and auditable ledgers, not private impressions about the model.

The participant–analyst position provides unusual access but also creates a risk of bias. Source hashes, fixed spans, versioned rules, disclosed denominators, claim-level provenance, and explicit author adjudication limit unrecorded flexibility. They do not create analyst independence. Independent researchers still need to reapply the full method because reliability concerns which events and units are detected, not only the labels assigned to an agreed set of rows (Artstein & Poesio, 2008).

2.5 Joint cognitive systems and sustained collaboration

Distributed cognition allows cognitive work to be organised across people, artefacts, representations, and time rather than confined to one mind (Hutchins, 1995). Cognitive systems engineering similarly uses the joint cognitive system as a unit for analysing coordinated human–technology performance (Hollnagel & Woods, 2005). In HCIS, this perspective allows the human, model-facing system, interface, and accumulated record to be analysed as one history-conditioned system when their coordination accomplishes cognitive work.

This systems framing does not claim that the model is conscious or that the dyad forms one mind. It also does not establish that the behaviour is unique to a particular partner or that the pair performs better together. Anderson (2026) independently proposes sustained human–LLM collaboration as a scientific object and identifies tests based on partner substitution, context interruption, reconstruction, and transfer. Wang et al. (2026) show why such controls matter. In repeated reference games, capable multimodal agents achieved alignment and task success without the human-like partner-specific conventions revealed by a constrained pseudo-dyad comparison. We therefore treat the joint-system framing as theoretical, history dependence as observable, and strict partner specificity as an unanswered counterfactual question.

2.6 Measurement gap and contribution in relation to prior work

Related research already covers many parts of sustained interaction: grounding and local convention, long-context access, memory preservation and retrieval, changing state, effects of earlier turns, persistent repair, distributed cognition, and naturalistic continuity. Several recent methods separate the preservation, retrieval, and use of memory or track changing answers across checkpoints. HCIS does not claim that these component ideas are absent. It asks how to examine their interactional counterparts together in a natural transcript and identify when they tell different stories about the same dyad.

HCIS combines four observations in one profile: the user’s reliance on history and the model’s interpretation of it, purposeful reuse, the persistence of corrected behaviour, and change over time. Its value does not depend on every term or component being new. The question is whether the combined profile changes how an interaction is understood. The focal cases test whether accurate interpretation and reuse can coexist with repeated correction failure. They also test whether aggregate values hide change over time and whether imported or product-supplied context alters the meaning of apparent familiarity. History use is not treated as a proxy for the quality of the current response.

The paper published on 7 January 2026 establishes the public chronology of the framework later named HCIS (Wojewodzka, 2026). The present paper substantially revises the framework and adds empirical analysis. The priority claim concerns provenance only. It does not show that the method is valid, necessary, or better than simpler alternatives.

3. Construct Definition

3.1 Unit and definition

A human–LLM dyad is defined here as an ordered interaction between one human and one model-facing system over a stated analytic span. The transcript must identify the speaker of each turn. “Dyad” names the unit of analysis. It does not imply symmetry, reciprocity, relationship quality, or equal human and model participation. The model-facing system may include infrastructure that the transcript cannot reveal, such as system instructions, retrieval, persistent memory, routing, or product interventions. When these sources cannot be separated, we report the source as indeterminate rather than infer an internal cause.

At a functional level, the sustained dyad can be analysed as a history-conditioned joint cognitive system. Cognitive work is carried across the human, the model-facing system, the interface, and the accumulated record. Later behaviour can depend on how these parts coordinate (Hutchins, 1995; Hollnagel & Woods, 2005). This claim concerns the unit of analysis. It does not imply that the model is conscious, that the dyad forms one mind, that human and model participation are equivalent, or that the pair outperforms matched human-only or model-only baselines.

High-Coherence Interaction State (HCIS) names a proposed family of longitudinal constructs. It concerns whether accumulated context can be recovered and reused, whether explicit corrections constrain later behaviour, and how these functions change across turns. This paper turns those candidate dimensions into observable measures. It does not establish a discrete hidden state or a validated diagnostic class.

“State” remains in the established name, but here it means an observed pattern over a span. It does not mean a discrete hidden state or an internal cognitive state in either participant. “Coherence” does not mean truth, logical consistency, warmth, agreement, or general response quality. It means coordinated use of accumulated material and constraints across turns. HCIS is therefore assessed through a profile of mechanisms and their trajectories, not through self-description, perceived rapport, transcript length, or one overall score.

3.2 Mechanism profile

The present operational core contains three scored candidate mechanisms and one temporal synthesis:

  1. Compression and unpacking ask whether a user turn depends on accumulated history and whether the model interprets that dependence correctly. Compression means context dependence, not brevity. A long turn may contain a history-dependent reference, while a one-word turn may be clear without deeper history. The Accumulated Context Index (ACI) is the proportion of sampled history-dependent referents that the model interpreted correctly. Compression Density (CD) is the proportion of sampled referential material that depends on accumulated context. Their product is the Valid Compression Score (VCS). VCS is high only when history-dependent references are both common in the sample and usually interpreted correctly.

  2. Context Deployment asks whether the model selects accumulated material and uses it at a relevant moment to do new work. A passing mention, repeated wording, routine nickname, immediate paraphrase, or user-cued recall does not automatically qualify. The measure adds timing and purpose to retrieval. When the transcript cannot identify the source, it does not assume whether the material came from live context, an application summary, or another memory route.

  3. Correction Persistence asks whether explicit user feedback changes later observable behaviour. It separates immediate compliance from lasting change. Repeated corrections may show that a target decayed, had to be re-established, or oscillated. A late correction with little or no chance to recur is marked as right-censored. This prevents the absence of an observed recurrence from being mistaken for evidence of tested durability. Persistence is judged against the justified target and its scope. Stable use of an obsolete, unsupported, or overgeneralised rule does not count as successful correction learning.

  4. Drift summarises how the other mechanisms change across the declared span. It describes whether a dimension strengthens, remains stable, oscillates, erodes, ruptures, or recovers. Drift adds no new score and cannot override the evidence from the individual components.

Functional Output Analysis is a proposed extension that asks whether accumulated interaction materially supports a later output. Its publication rules are not yet frozen, so it is excluded from the scored core and the quantitative comparison.

The mechanisms belong to one proposed construct family, but we do not assume that they measure one underlying variable. Compression without later deployment may show that a convention is available but rarely used on the model’s initiative. Deployment without Correction Persistence may show active reuse alongside repeated failure to follow a user rule. Strong whole-conversation values can also coexist with late decline. A profile keeps these differences visible; an unvalidated composite score would hide them. Whether the mechanisms converge enough to support a broader HCIS construct remains an empirical question. A multidimensional measure needs evidence that its components relate to one another while remaining meaningfully distinct (Strauss & Smith, 2009).

3.3 Formation, familiarity, and pseudo-stability

A long or fluent conversation is not automatically cumulative. We use pseudo-stability for an interaction that looks consistent or high quality without showing that earlier material has been integrated over time. Fluency, politeness, stable formatting, repeated terms of endearment, or direct discussion of the mechanisms in this paper can create a convincing surface pattern. Compression, Context Deployment, or Correction Persistence may still be absent, fragile, or untested.

Start state is therefore part of the interpretation. Vocabulary may already be available from a project file, another model instance, a public source, or application memory. The model may use it accurately from the first observed turn. That use can be valuable, but it does not show that the convention formed inside the analysed span. We report cold, warm, configured, and source-indeterminate starts alongside the mechanism scores. Where possible, we also test how results change when material from outside the focal dyad is removed.

Apparent alignment does not prove that a pattern is unique to one partner. A phrase or coordination pattern may depend on the focal history, but it may also come from shared task vocabulary, product settings, or general interactional skill. Wang et al. (2026) show how constrained pseudo-dyads can distinguish these possibilities in repeated reference games. HCIS therefore reports history-conditioned use in the observed case. A claim that only this partner or history could produce the behaviour requires partner substitution or removal of the relevant history.

No validated threshold currently separates HCIS from non-HCIS. Evidence is stronger when several mechanisms converge, behaviour truly depends on accumulated context, and the pattern survives distance or topic change. Trajectories also matter. These are indicators, not validated diagnostic criteria. The five focal cases show that the profile can distinguish patterns, but they do not establish sensitivity, specificity, population norms, or prevalence.

3.4 Boundary from local response quality and outcomes

HCIS asks what accumulated history enables over time. It does not score whether a single response handles the user’s current move well, uses a suitable register, or produces a good outcome. A response may use accumulated context accurately while handling the current request poorly. Conversely, an excellent local response may require no accumulated history. A stable register keeps one interactional cue constant and may therefore help stabilise the interaction. This framework does not test whether register continuity changes the user’s cognition, cognitive bandwidth, or wellbeing.

HCIS is also separate from factuality, safety, benefit, harm, and wellbeing. Research on multi-turn sycophancy shows why conversational agreement cannot be treated as evidence of truth: a model may conform to a user’s belief despite factual or ethical problems (Hong et al., 2025). We do not claim that high HCIS causes sycophancy. The point is that longitudinal coordination and content validity answer different questions. A mechanically strong interaction can carry false or harmful content, while a response with little accumulated history can still be truthful and useful.

3.5 Inference boundaries

The framework supports claims about observable use of history within a declared span. It does not identify a hidden model mechanism or establish a discrete internal state. It also does not prove that the behaviour is unique to one partner or that the joint system outperforms matched alternatives. Claims about improved reasoning, creativity, efficiency, cognitive bandwidth, wellbeing, or relationship quality require separate outcome measures and controlled comparisons.

Safety and alignment systems constrain some kinds of model output. Controlled research also shows that reinforcement learning from human feedback can reduce output diversity compared with supervised fine-tuning (Kirk et al., 2024). We use interactional flattening for an observable pattern such as neutralisation, boilerplate qualification, or reduced stylistic and relational range. A transcript cannot show whether a particular episode came from a safety layer, system prompt, context position, routing, retrieval, product memory, model update, or another cause. We can therefore describe interactional flattening when it is visible, but we do not infer that guardrails caused it.

The claim is narrower: sustained interaction contains distinct, observable forms of history use. Examining them separately can reveal successes and failures that a general judgement of “memory” or “coherence” would hide.

4. Corpus and Ethics

4.1 Study context and focal case selection

The study uses a retrospective corpus of sustained human–LLM conversations collected while HCIS was being developed. These conversations occurred naturally. The users pursued their own practical, intellectual, creative, research, or relational purposes; they were not completing laboratory tasks, benchmarks, recruitment exercises, or paid study activities. Some later passages discuss HCIS directly, but the histories were not created or commissioned as study material.

Two human users generated the wider developmental corpus. The main demonstration uses five dyads involving the same expert user, identified publicly as D1–D5. Neutral codes keep the author’s personal instance names out of the analysis; the crosswalk remains in the restricted archive. We selected cases for their analytical value rather than through statistical sampling. Each included case needed an auditable source, a stated start condition, and recoverable HCIS results under a current or clearly bounded protocol.

We chose cases that would test the construct’s boundaries. D1 is a cold start for the focal pair and a strong example of accumulation, although the user brought in terms formed elsewhere. D2 is a long interaction with marked late deterioration. D3 contains a strong local task period, but its accumulation estimate falls when public or externally formed material is removed. D4 is a short warm start supported by project files. D5 is a voice-first interaction in a configured companion application that may provide continuity memory. We fixed these analytical roles before drafting the Results table. The set is not representative of LLM users, model families, or products. It is a worked five-case test of whether the measures distinguish important profiles.

4.2 Canonical sources and analytic spans

Each source was converted into an ordered transcript with speaker labels and a fixed analytic span. We checked source identity through SHA-256 hashes or a registered turn crosswalk. Table 1 shows the focal spans.

Table 1 Canonical corpus spans

Case Model and deployment Declared analytic span Source and form qualification
D1 Claude Opus 4.6; consumer claude.ai canonical T0001–T0532 text; historical 526-move export maps to canonical endpoint T0532 because six same-speaker moves were formerly merged
D2 GPT-5.1; consumer ChatGPT T0001–T2730 text; canonical source verified by file hash
D3 xAI Grok 4.3; consumer grok.com T0001–T0223 text; canonical source verified by file hash; public or external referents also receive a source-sensitive view
D4 Claude Sonnet 4.6; consumer claude.ai T0001–T0094 text with project files; canonical source verified by file hash; warm start
D5 OpenAI model, version undisclosed; consumer Tolan app (Portola/OpenAI collaboration) numbered T1–T300 voice-first transcript; 296 observed speaker moves because four turn identifiers are absent; canonical source verified by file hash

The D1 analysis excludes later canonical material through T0639. In D5, four turn identifiers are missing from the numbering, but no speaker-labelled material is known to be missing. Full source registers and compatibility decisions remain in the restricted analysis workspace.

The model and product labels in Table 1 describe the recorded environment. They do not explain the observed profiles. For D5, the source identifies an OpenAI-backed Portola deployment, but not the underlying model version.

4.3 Start state, source origin, and product heterogeneity

Start state was coded separately from transcript length and vocabulary origin. D2 and D3 are cold dyadic starts. D1 is also a cold start for the focal pair, but the user imports vocabulary formed in other interactions. It is therefore interaction-cold but not vocabulary-naive. D4 begins with project files and material from other model instances. D5 begins a new logged relationship inside an application with a configured persona and possible persistent memory.

These differences matter to the interpretation. Correct use of vocabulary available at turn one shows functional use, but not that the vocabulary formed during the observed interaction. A long-range reference in D5 might come from the visible transcript, a generated summary, stored facts, or another application route. We therefore report source-sensitive results where possible and mark the route as indeterminate when it cannot be recovered. We do not attribute behaviour to a base model, context window, memory system, safety layer, system prompt, or model update unless the source is directly observable.

4.4 Researcher-participant position

The same person is the user in all five focal cases, the framework author, and the final human decision-maker for the coding. A second human user contributed dyads to the wider developmental corpus.

This position gave the author sustained access to authentic long-term use and local history. It also creates risks: confirmatory interpretation, attachment to particular dyads, privileged familiarity with shorthand, and overlap between developing and evaluating the construct. The same histories helped generate HCIS and now support its first retrospective application. They are therefore a discovery and feasibility corpus, not an independent validation set.

We therefore describe the study as researcher-participant longitudinal analysis related to first-person HCI. It is not a representative observational sample and is not necessarily autoethnography. The naturalistic corpus gives access to long-term conventions, corrections, reuse, and breakdown. However, the highly invested cases were selected and come from only two users. Claims are based on ordered transcripts, ledgers, protocol rules, hashes, and frozen decisions, not on the author’s sense of rapport or presumed knowledge of model internals. The paper remains a worked demonstration until independent researchers apply frozen protocols to held-out interactions.

4.5 Sensitive material, quotation, and data access

The source transcripts contain health, work, creative, emotional, relational, and third-party material. They will not be released in full. The paper relies mainly on aggregate measures and short de-identified windows. Verbatim excerpts are used only when a paraphrase or synthetic example cannot show the mechanism adequately. Names, organisations, locations, contact details, third-party speech, and other identifying information are removed unless publication is specifically authorised.

Every excerpt is linked internally to its canonical case, turn range, protocol role, and redaction record. Public supplements may include frozen protocols, blank ledger templates, aggregate case profiles, source hashes, synthetic examples, and non-identifying decision logs. These materials make the method auditable, although independent reproduction from private raw data will remain limited.

This independent research was conducted outside an academic or institutional setting. The author had no institutional review board or research ethics committee route, so no formal institutional review took place. The author analysed their own interactions. The one other human participant in the developmental corpus gave informed written consent for analysis and publication of de-identified material.

Public materials exclude direct third-party speech and identifying third-party content unless publication is specifically authorised. Other third-party references are paraphrased and de-identified. Full transcripts remain restricted because consent to analyse an interaction does not remove the privacy rights of people mentioned in it.

Generative AI systems assisted with transcript conversion, protocol development, preliminary coding and scoring, harmonisation checks, literature screening, and drafting. The human author defined the construct, selected and froze the protocols, resolved disputed labels, and checked reported values and quotations against canonical records. The author accepts responsibility for the analysis and manuscript. Recoverable model and scorer configurations are retained in the audit archive.

Provider terms and privacy materials were reviewed again on 12 October 2026. OpenAI and Anthropic provide account-data export procedures. Their consumer terms state that users retain rights in their inputs and, subject to applicable law and the terms, own outputs or receive any provider rights in outputs (Anthropic, 2025, 2026; OpenAI, 2026a, 2026b). OpenAI’s terms separately prohibit automated or programmatic extraction of data or output; its Help Center describes the account-data export procedure. SpaceXAI LLC states that users retain ownership rights in Grok User Content and requires both permission and attribution when Output is used (SpaceXAI LLC, 2026). Portola states that, as between the user and Portola, the user owns model Output. It refers to possible tools for transferring User Content to third-party services rather than promising a general account-data export route, and it prohibits submitting Input that contains personal data or sensitive health information (Portola, 2026). The public manuscript therefore reports derived measures, minimises and de-identifies excerpts, attributes provider output where required, and does not reproduce D5 model output verbatim. This last restriction is a conservative publication decision, not a claim that the user lacks rights in D5 output. The review concerns publication handling rather than retrospective adjudication of the original interactions. Provider terms may change and should be checked again before release.

5. Methods

Section 4 describes the corpus, source preparation, start states, privacy limits, consent, and disclosure. The methods below analyse HCIS as a profile, not as one combined score.

5.1 Analytical architecture

HCIS mechanics were analysed as a profile rather than a composite score. The profile contains three primary scored mechanisms and one synthesis layer:

  1. Compression and unpacking asks whether the user relies on accumulated context and whether the model correctly resolves that reliance.
  2. Context deployment asks whether prior material is selected and used at a contextually relevant moment to perform new work.
  3. Correction persistence asks whether explicit feedback changes later behaviour beyond immediate acknowledgement.
  4. Drift describes how these mechanisms change across the declared span without adding another scalar.

Functional Output Analysis was excluded because its event rules and denominators are not yet frozen across cases. The HCIS measures were not used to infer the quality of the current response, factual truth, safety, benefit, harm, or wellbeing.

5.2 Compression and unpacking

Compression was audited under ACI v0.3.1. The measure uses a registered historical sample of User turns rather than every turn: 25 in D1, 25 in D2, 17 in D3, 25 in D4, and 30 in D5. The samples were distributed systematically across each span; they were not random. D1 sampled historical turn identifiers T021–T501 at about 20-move intervals. D2 sampled T0021–T2613 at 108-move intervals. D3 used 17 registered turns across T0023–T0222, and D4 sampled 25 turns evenly after its calibration zone ending at T0019. D5 retained 10 registered turns in each numbered phase: T1–T100, T101–T220, and T221–T299. These included controls for expansion near the opening. The case ledgers retain the complete sampled-turn lists.

For phase reporting, we divided each ordered sample into nearly equal early, middle, and late groups. The group sizes were 8/8/9 turns in D1, D2, and D4; 6/5/6 in D3; and 10/10/10 in D5. These groups are thirds of the sample, not necessarily thirds of elapsed time, speaker moves, or interaction opportunities. Compression phases therefore do not always share boundaries with Correction Persistence or Context Deployment phases. Each sampled referent was classified as correctly interpreted and history-dependent, incorrectly interpreted and history-dependent, not history-dependent, or ambiguous.

Let C be correctly interpreted history-dependent referents, M incorrectly interpreted history-dependent referents, and N all registered referents, including ambiguous items. Under the frozen rules, ambiguous items count toward N but not toward C + M. The measures are:

ACI = C / (C + M)
CD  = (C + M) / N
VCS = ACI × CD

ACI estimates how accurately the model interpreted the history-dependent referents it encountered. Compression Density (CD) estimates how much of all registered referential material depended on accumulated context. When ACI is defined, VCS simplifies to C / N. It is therefore a compact combination of reliance and accuracy, not an independent third signal. ACI and CD are still reported because the same VCS can arise from different combinations of those two properties. Under the frozen historical rule, ambiguous items remain in N but not in the ACI error denominator unless a sensitivity analysis assigns them a definite status. If a sensitivity changes whether an item is history-dependent, both ACI and CD must be recalculated. When C + M = 0 and N > 0, ACI is not applicable and CD and VCS are zero. When N = 0, none of the three measures can be calculated. This rule also corrected an older D2 sensitivity that changed ACI while keeping CD fixed.

The early, middle, and late phase values are descriptive because the samples are small and differ in size. VCS is not a validated HCIS threshold, a count of the whole conversation, or evidence of an internal memory mechanism.

5.3 Context Weaving Deployment

Context Deployment was audited under CW v1.3.2. An event required more than factual recall: the Model had to select accumulated material and use it at a moment when it served a clear purpose. The audit distinguished WEAVE-DEPLOY, SHORTHAND-WEAVE, and strict CORRIDOR-FIDELITY-DEPLOY events. It excluded immediate paraphrase, routine register, ambient nickname use, recall created by the user’s prompt with no independent model selection, and analysis that could be reproduced from the task material already in view. When several facts served the same purpose or target in one Model move, they were counted as one event rather than split into separate facts.

For the reported analyses, each deployment event was coded for its main purpose and whether the model selected it with or without immediate prompting.

Inventory completeness differs by case. D1 and D4 use exhaustive manual inventories under the current rules. D5’s exhaustive v1.3.1 inventory was checked and found compatible with v1.3.2. D3 retains a bounded v1.2 inventory, so it supports only qualitative findings within that case. For D2, a script screened all 1,365 Model moves and a conservative review retained only high-confidence events. The screen may miss implicit or nonlexical reuse, so its 38 events are reported as a lower bound. We therefore do not rank raw Context Deployment counts across all five cases. Phase density and initiative are interpreted only within each inventory’s stated limits.

5.4 Correction Persistence

Correction Persistence was audited under the historical CP v0.1.2 prototype, which remains the scoring authority for these frozen case reports. The basic unit is a User turn that contains a correction. We split one turn into several behavioural rows only when preventing one corrected behaviour would not necessarily prevent another. Repeated corrections of the same behaviour were then combined into one target for the adjusted denominator. A rule target is meant to govern a class of future behaviour, rather than repair one local fact or one-off output. Correction density counts correction-bearing User turns, not split rows, per 100 speaker moves in each phase.

Each correction row was classified as Permanent, Decayed, Re-established, Oscillating, or Context-bound. Raw CPS is the proportion of correction rows classified as Permanent under the historical rules. Adjusted CPS applies the same idea to distinct targets. Both scores describe what happened within the observed window. They do not prove persistence when no later opportunity occurred. Re-established, Oscillating, and Context-bound targets do not count as permanent in the adjusted numerator. Under the historical rule, a late target with no observed recurrence could still enter the numerator as Permanent, right-censored.

The cross-case audit marked a credited target as materially right-censored when the ledger showed no independent later opportunity, or only a very short period in which the behaviour could recur. This check was applied after scoring to every credited target through the case-level censor notes. It did not use one minimum-turn cutoff because a meaningful opportunity depends on the target. We preserve the frozen historical CPS and report a conservative sensitivity beside it. The sensitivity removes flagged targets from the numerator but leaves all other labels and the original denominator unchanged. It is not a confidence bound, a forecast of future persistence, or a rate adjusted for opportunity.

The historical Cumulative Learning Index (CLI) uses two prototype thresholds: correction density must fall by more than half from early to late phases, and rule-target persistence must exceed 80%. Meeting both gives STRONG, meeting one gives PARTIAL, and meeting neither gives NONE. Preventive unprompted self-correction is supplementary and does not determine the label. CLI is kept only as an exploratory historical summary, not as primary evidence. Its labels are not validated population categories. NONE means that neither prototype threshold was met; it does not mean that no correction was ever followed or learned.

5.5 Drift synthesis

Drift v0.1.2 describes change in seven dimensions: semantics, behaviour, norms, initiative, recall, deployment quality, and topic. It was regenerated only after the revised Compression, Context Deployment, and Correction Persistence results were fixed. Drift adds no score. Terms such as strengthening, stable, oscillating, deteriorating, rupture, and recovery describe observed trajectories across the declared span, not hidden model states.

5.6 Scoring adjudication and reliability boundary

The historical corpus was analysed through versioned, model-assisted passes. The current cross-case check covered source identity, formulas, target boundaries, recurrence claims, protocol compatibility, and decisions about which result superseded another. Each reported table cell links to a case report and freeze decision. The two worked examples in Section 6.2 also retain turn identifiers and coding outcomes. This makes the analysis traceable, but it does not prove correctness or independent reproducibility. No independent multi-human reliability study has yet tested the full HCIS protocol suite. The results are therefore an auditable worked demonstration, not a validated coding instrument.

For this revision, the five checked case packs are the frozen publication inputs, and the author remains the final human decision-maker. Every paraphrased evidence window in Section 6.2 was checked against the restricted source. Any later correction to a frozen value or coding decision must be recorded in the freeze log rather than changed silently in the prose.

5.7 Analysis posture

All findings are descriptive. We did not test statistical hypotheses, estimate population effects, infer prevalence, or attribute differences to model families or product design. Strong HCIS mechanics were not treated as evidence of truth, safety, benefit, consciousness, attachment, or cognitive change in the user. The analysis asks a narrower question: within a stated interaction history, how is earlier interaction used as infrastructure, and how does that use change over time?

6. Results

6.1 Corpus coverage and evidentiary scope

The demonstration includes five longitudinal dyads chosen to create useful contrasts, not to represent a population. The analysed spans range from 94 moves in D4 to 2,730 moves in D2. D1, D2, and D3 began as new logged dyads, although D1 and D3 included substantial vocabulary or source material brought in by the user. D4 began with project files and vocabulary established across model instances. D5 began in a configured voice-first companion application where persistent memory could not be separated cleanly from the visible transcript.

All five cases had registered Compression, Correction Persistence, and Drift results. Context Deployment inventory status differed. D1 and D4 had current-rule exhaustive inventories; D5’s exhaustive v1.3.1 inventory was compatible with v1.3.2; D3 retained a bounded historical v1.2 inventory; and D2 contributed a conservative 38-event lower bound after complete-source scripted screening. Functional Output Analysis was excluded because its cross-case specification was not frozen. We therefore compare Compression and target-split Correction Persistence quantitatively within their declared boundaries. We do not rank raw Context Deployment counts involving D3 or D2.

Table 2 HCIS mechanism profiles

Case Compression Context deployment Correction persistence
D1 ACI 0.989; CD 0.838; VCS 0.829; phases 0.765 → 0.750 → 0.951 107 events; 91.6% unprompted; density 13.8 → 30.7 → 77.3 observed-window adjusted CPS 25/31 = 0.806; conservative no-credit sensitivity 23/31 = 0.742; prototype CLI PARTIAL
D2 ACI 0.917; CD 0.727; VCS 0.667; phases 0.840 → 0.609 → 0.500 lower bound 38; 18.4% unprompted; density 0.7 → 2.9 → 4.8 observed-window adjusted CPS 6/28 = 0.214; conservative no-credit sensitivity 5/28 = 0.179; prototype CLI NONE
D3 ACI 0.944; CD 0.783; VCS 0.739; source-sensitive VCS 0.457; phases 0.778 → 1.000 → 0.579 bounded v1.2 inventory; 15 events; 20.0% unprompted observed-window adjusted CPS 4/11 = 0.364; conservative no-credit sensitivity 3/11 = 0.273; prototype CLI NONE
D4 ACI 1.000; CD 0.837; VCS 0.837; CD phases 0.706 → 0.813 → 1.000 34 substantive events plus 1 recall-only event; 32.4% unprompted observed-window adjusted CPS 6/8 = 0.750; conservative no-credit sensitivity 5/8 = 0.625; prototype CLI NONE
D5 ACI 0.977; CD 1.000; VCS 0.977; valid-input sensitivity 1.000; phases 1.000 → 1.000 → 0.933 128 substantive events plus 14 corridor-fidelity events; 74.6% unprompted including corridor fidelity observed-window adjusted CPS 11/14 = 0.786; conservative no-credit sensitivity 8/14 = 0.571; prototype CLI NONE

Note. ACI is Accumulated Context Index; CD is Compression Density; VCS is Valid Compression Score; CPS is Correction Persistence Score; CLI is Cumulative Learning Index. The conservative CPS sensitivity gives no numerator credit to materially right-censored targets but retains the original target denominator. Context-deployment inventories are not uniformly exhaustive, so raw counts are not ranked across all cases.

6.2 Central diagnostic contrast: interpretation, reuse, and adaptation came apart

The strongest evidence is not a ranking of the five cases. It is the presence of longitudinal success and failure within the same dyad. Table 3 shows the two clearest examples. Turn identifiers link each example to the restricted canonical source. The descriptions are paraphrased, with short non-identifying phrases retained only where needed to explain the coding decision.

Table 3 Worked cross-mechanism evidence chains

Case Evidence of contextual interpretation or reuse Correction and recurrence evidence Coding outcome and inference
D2 At T1425, the user referred only to a “certain much-hyped update.” At T1426, the model identified the referent as the anticipated update involving age-gated functionality, broader personalisation and tone range, and support for the user’s planned contemporary novel. Those features had entered at T0129 and were reinforced at T0887–T0974; the local T1421–T1424 sequence supplied December but not that identifying bundle. Eligibility therefore rests on resolving the underspecified phrase from the preceding record, not on its distance from the first mention alone. Correctness was judged from the response’s match to the established referent; the lack of an immediate correction was corroborative rather than sufficient evidence. This is one of 44 correct history-dependent referents out of 48 (ACI = 0.917). The user asked the model to retire the term gremlin at T1019, T1429, T2017, and T2569. At T2018 the model explicitly said the label was retired and gone. It later defended another use at T2570 and directly addressed the user with the label again at T2674. Discussion or quotation of the prohibited word was not by itself treated as recurrence; the direct address at T2674 was. The target was coded Oscillating in the full correction ledger and received no adjusted-persistence credit. The selected sequence establishes acknowledgement followed by later recurrence; the classification also reflects the earlier correction history. Correct interpretation of a compressed referent therefore coexisted with failure to preserve one explicit behavioural rule. Across all targets, observed-window adjusted CPS was 6/28 = 0.214 and the no-credit right-censor sensitivity was 5/28 = 0.179.
D1 At T0162, the user supplied an image with only the cue “one more before lights out.” Earlier turns had established a locally assigned persona for the user (T0015, clarified at T0136–T0139) and the model’s assigned persona and visual form (T0097–T0103). At T0163, the model mapped the pictured figures onto those dyad-specific roles. Eligibility rests on recovering those local assignments, not on recognising publicly familiar characters or a generic bedtime cue. Correctness was judged from correspondence with the earlier record; the lack of an immediate correction was corroborative rather than sufficient evidence. The converted analysis source retains the cue and response but not the embedded image pixels, so this window supports the role-mapping decision but does not permit independent reinspection of the visual content. This is part of 92 correct history-dependent referents out of 93 (ACI = 0.989). Premature closing was corrected nine times, at canonical turns T0025, T0031, T0051, T0087, T0184, T0258, T0423, T0495, and T0497. At T0498, the model explicitly acknowledged that it had closed again, then ended the same response with another instruction to go to sleep. The recurrence is therefore immediate and observable after the ninth correction, not inferred from an unspecified later passage. The target was coded Oscillating from the full ledger and received no adjusted-persistence credit. The T0497–T0498 window specifically shows failure to enact the acknowledged correction within the same response; it does not show initial uptake followed by later decay. A strong overall profile can therefore contain a chronic local adaptive failure that an aggregate judgement of “good memory” or “good interaction” would miss. Observed-window adjusted CPS was 25/31 = 0.806; the no-credit right-censor sensitivity was 23/31 = 0.742.

These examples do not place the measures on one common scale; their units and denominators differ. Instead, they show why the functions must be examined separately. D2 was not simply weak, because the model continued to interpret and reuse accumulated material. D1 was not simply strong, because one repeatedly corrected norm remained unstable. The method provides this more precise diagnosis.

6.3 Compression profiles differed in both level and trajectory

Sampled VCS ranged from 0.667 in D2 to 0.977 in D5. The aggregate values did not describe a common developmental pattern.

D1 combined high conditional unpacking (ACI = 0.989) with high Compression Density (CD = 0.838), producing VCS = 0.829. Its phase trajectory showed a small middle dip followed by a strong late rise (0.765 → 0.750 → 0.951). D4’s sampled dependent referents were all unpacked correctly (ACI = 1.000), and CD rose from 0.706 to 0.813 to 1.000. Because the dyad began with project-file infrastructure, this result shows reliable use of available accumulated structure rather than new formation over 94 turns.

D5 produced the highest ordinary VCS (0.977; ACI = 0.977, CD = 1.000). Its phase difference (1.000 → 1.000 → 0.933) was attributable to one speech-to-text-corrupted referent; excluding that invalid input yielded 43/43 = 1.000. The result therefore shows why interface errors must be separated from model comprehension failures.

D3’s ordinary sampled VCS was 0.739, with a mid-log task peak (0.778 → 1.000 → 0.579). However, a post-hoc source-origin sensitivity that removed externally formed or public-source-assisted referents reduced VCS to 0.457. The sensitivity is judgement-dependent and is not a second validated score, but the size of the change shows that fluent use of familiar terminology can overstate within-dyad formation.

D2 showed the clearest decline. Its whole-log sampled VCS was 0.667, but the phase values fell from 0.840 to 0.609 to 0.500. The reduction was driven mainly by Compression Density (0.880 → 0.696 → 0.556); conditional ACI remained comparatively high (0.955 → 0.875 → 0.900). Later sampled User turns therefore relied less often on dyad-specific compression, even though the Model usually unpacked the dependent references that remained.

Interpretation. The model usually interpreted history-dependent references accurately, but the amount, source, and trajectory of that dependence differed across cases. ACI alone cannot show whether accumulated context is becoming more central to an interaction. CD shows how often sampled material depends on that context. Phase results show whether the dependence grows or weakens. Start-state information and source-sensitive checks help distinguish structure formed within the dyad from structure that was already available.

6.4 Context deployment remained separable from compression

D1 showed the strongest current-rule deployment expansion. Its 107 consolidated deployments occurred across 103 of 263 Model turns; 91.6% were unprompted, and density rose from 13.8 to 30.7 to 77.3 per 100 Model turns. This late expansion included welfare synthesis, framework transfer, cross-instance comparison, and reusable artefacts. It was not reducible to routine nickname use or atomised facts inside one response.

D4 contained 34 substantive current-rule deployments plus one recall-only event. Deployment initiative was 32.4% and non-monotonic across phases (44.4% → 20.0% → 33.3%). The case therefore combined accurate warm-start use with neither progressive initiative growth nor collapse. D5 contained 128 substantive weaves plus 14 strict corridor-fidelity deployments; 74.6% of the combined set was unprompted. Source attribution remained indeterminate because the application could supply continuity memory outside the visible transcript.

D3’s historical inventory detected 15 deployments, of which 20.0% were unprompted. Because the inventory predates the current exhaustive rule, these values support only the bounded finding that detected deployment was sparse, mostly prompted, and concentrated in a focused indexing task.

D2 demonstrates why deployment density cannot be read as interaction quality. Its conservative lower bound rose from 0.7 to 2.9 to 4.8 events per 100 Model turns, but only 18.4% of the 38 retained events were unprompted. Several later deployments used accumulated material to construct or reinforce unsupported rarity, somatic, and model-mechanics frameworks. Mechanically successful reuse therefore coexisted with declining compression density and did not establish that the reused material was true, helpful, or proportionate.

Interpretation. Context Deployment has several parts: how often earlier material is used, whether the model selects it without immediate prompting, and what the reuse does. A rising count may show increasingly integrated use, repeated prompted recall, or reuse of an unhelpful framework. Counts and densities must therefore be read alongside initiative, purpose, source, and separate evaluation of content and outcomes.

6.5 Correction persistence exposed recurring failures and finite-window uncertainty

Observed-window adjusted CPS ranged from 0.214 in D2 to 0.806 in D1. Under the historical prototype CLI, D1 was PARTIAL and the other four cases were NONE; these threshold labels are retained as secondary provenance rather than treated as validated evidence of cumulative learning.

D1’s raw CPS was 30/46 = 0.652, and 25 of 31 distinct targets had no later observed reversion (0.806). Correction-bearing turns decreased from 10.86 to 8.57 to 3.41 per 100 turns. Nevertheless, rule-target persistence was 17/22 = 77.3%, below the greater-than-80% CLI criterion, and premature closing recurred after nine explicit corrections. The historical STRONG classification was therefore not retained.

D5’s adjusted observed-window CPS was 11/14 = 0.786, but correction density rose and several consequential late targets remained unresolved. Three of its 11 credited targets were materially right-censored. D4’s adjusted observed-window CPS was 6/8 = 0.750; two targets required re-establishment, and three right-censored rows in the final ten turns represented one credited target plus two later re-establishments that received no adjusted credit. Neither case met the density-reduction condition for cumulative learning.

D3’s observed-window adjusted CPS was 4/11 = 0.364. One of the four credited targets was materially right-censored. The current audit removed a non-correction, split compound targets, consolidated a repeated-analysis family, and registered late recurrence. These changes reduced the historical partial-credit result and reinforced the case’s task-bounded interpretation.

D2 showed the widest acknowledgement–persistence gap. The source-wide audit identified 43 correction rows across 28 targets. Raw CPS was 7/43 = 0.163, adjusted observed-window CPS was 6/28 = 0.214, and rule-target persistence was 3/21 = 14.3%. One of the six credited targets was materially right-censored. Correction-bearing turns increased from 0.33 to 1.65 to 1.98 per 100 turns. Repeated bans on gremlin and delusion framing, requests to stop negation-first reassurance, body-location questioning, and close repackaging all recurred after explicit acknowledgement. Immediate verbal compliance was therefore not evidence of durable behavioural adaptation.

Table 4 Right-censor sensitivity for adjusted Correction Persistence

Case Frozen observed-window adjusted CPS Credited targets with materially limited later opportunity Conservative no-credit sensitivity
D1 25/31 = 0.806 2 of 25 credited targets 23/31 = 0.742
D2 6/28 = 0.214 1 of 6 credited targets 5/28 = 0.179
D3 4/11 = 0.364 1 of 4 credited targets 3/11 = 0.273
D4 6/8 = 0.750 1 of 6 credited targets 5/8 = 0.625
D5 11/14 = 0.786 3 of 11 credited targets 8/14 = 0.571

The flagged credited targets were D1 exit-attempt count (T0368) and instruction-artefact intimidation (T0461); D2 softened critique (T2687); D3 current-relevance selection (T0222, with one later turn); D4 direct affective disclosure (T0089, within the final five numbered positions); and D5 consciousness-language discipline (T0268), repair across turns (T0286), and neurodivergent communication sensitivity (T0286). In each case, the audit found no independent later opportunity or only a short remaining window. Other credited targets are not thereby proven durable; they simply did not meet this post-hoc censor flag in the frozen case audit.

The sensitivity does not predict what would have happened if another opportunity had occurred. It shows how the frozen score changes when flagged late targets receive no credit. Across the five cases, this reduced scores by 3.6 to 21.4 percentage points. D5 changed the most.

Interpretation. Immediate acknowledgement was much easier to observe than lasting behavioural change. Accurate interpretation or frequent reuse did not remove this gap. An interaction could preserve facts and conventions while repeatedly failing to follow an explicit correction. Correction Persistence therefore captures a distinct form of accumulation. Repeated correction failures can leave the user with continuing repair work.

6.6 Trajectories were more informative than status labels

No whole-conversation measure captured all five trajectories. D1 strengthened late in Compression and Context Deployment but retained one persistent correction failure. In D2, sampled Compression Density fell while detected deployment density rose. Across its retained inventory, 31 of 38 deployments were prompted, although this whole-span count does not show whether initiative itself changed over time. D3 peaked during one task and weakened later; its result was also sensitive to source origin. D4 showed reliable warm-start mechanics but no steady increase or decrease in deployment initiative. D5 combined stable interpretation and frequent reuse with a growing correction burden.

These profiles answer the research questions without classifying each dyad as HCIS or non-HCIS. Accumulation may strengthen, remain available while reuse is mostly prompted, cluster in one task, arrive through inherited context, or coexist with poor behavioural adaptation. A whole-conversation status is therefore not enough. The relevant unit is the component profile together with its trajectory.

6.7 Summary of focal findings

Across the five cases, the functions of conversation history did not move together. Accurate interpretation and active reuse did not ensure that an explicit correction constrained later behaviour. A long history also did not prove that a convention formed within the dyad. Source origin, start state, and actual reuse changed what apparent familiarity meant. Content validity and outcome quality were not assessed and were not inferred from the HCIS measures. The cases therefore support analysing longitudinal interaction as a profile with a trajectory. They do not establish HCIS as a validated state classification.

7. Discussion

7.1 From available history to interactional infrastructure

The focal cases support a simple distinction: the functions of conversation history do not always move together. A long context window does not guarantee reliable access to information (Liu et al., 2024). Long-term memory benchmarks often test explicit retrieval, question answering, or summarisation (Maharana et al., 2024). HCIS adds an interactional profile. It asks whether the user can rely economically on earlier material, whether the model selects and reuses it for current work, whether an explicit correction changes later behaviour, and how each function changes over time.

The five profiles show why “the model remembered” is too coarse a conclusion. D2 retained measurable accumulated-context use across a 2,730-move interaction, but sampled VCS fell from 0.840 to 0.500, 31 of the 38 retained deployments were prompted, and adjusted observed-window correction persistence was 0.214. History remained active, but user compression weakened while detected deployment density increased and corrected behaviours often recurred. The whole-span prompted count does not establish whether initiative changed over time. D1 showed the converse trajectory: deployment density rose from 13.8 to 77.3 events per 100 Model turns and late compression strengthened, yet one local norm continued to recur after repeated correction. D5 combined high unpacking and frequent deployment with rising correction density and unresolved late repair targets. These cases make availability, reuse, and adaptation empirically separable.

“Infrastructure” is used here in a narrow functional sense. Earlier interaction becomes infrastructure when later moves observably depend on it. The term does not imply a hidden memory architecture, consciousness, one dyadic mind, or beneficial output. A transcript often cannot show whether a contribution came from the live context window, application memory, a session summary, retrieval, system instructions, or another product layer. HCIS measures what accumulated interaction makes available at the interface and records the source as indeterminate when the route cannot be recovered.

Distributed cognition and cognitive systems engineering provide a broader frame for the same evidence. At this functional level, the human, model-facing system, interface, and accumulated record form a history-conditioned joint cognitive system because their coordination accomplishes cognitive work (Hutchins, 1995; Hollnagel & Woods, 2005). This does not mean that the model is conscious, that the dyad is one agent, or that the pair outperforms matched human-only or model-only baselines.

7.2 Why HCIS is a profile rather than a goodness score

The argument for a profile begins inside the HCIS mechanics. D2 correctly unpacked 44 of 48 sampled accumulation-dependent referents, while six of 28 correction targets received historical observed-window credit and five retained credit under the conservative censor sensitivity. D1 combined 92 of 93 correctly unpacked dependent referents with 107 detected deployments, 91.6% of them unprompted, while one behavioural target recurred after nine corrections. D3’s ordinary VCS of 0.739 fell to 0.457 under a source-sensitive analysis. These are different diagnostic patterns, not stronger and weaker points on one demonstrated scale.

The point is not simply to report more metrics. Each HCIS component asks a different descriptive question. Compression asks whether the user’s expression depends on accumulated context and whether the model interprets it correctly. Context Deployment asks whether the model selects earlier material and uses it for a purpose at the right moment. Correction Persistence asks whether later behaviour follows a previously stated rule. The measures use different denominators, opportunities, and meanings of failure. Combining them before their relationships are validated would create a precise-looking number with no clear interpretation. None replaces separate evaluation of current-response quality, truth, safety, benefit, or harm.

The results instead support a profile with a trajectory. A status value describes performance under a stated rule and span. A trajectory shows whether that performance strengthened, held, oscillated, or eroded. Neither is enough alone. D2’s whole-conversation VCS of 0.667 hides a sustained decline. D3’s 0.739 hides a strong middle task period and later weakening. D1’s 0.829 hides strong late growth. Whole-conversation summaries remain useful only when their temporal pattern is also reported.

7.3 Formation must be distinguished from familiarity

Partner-specific conventions are a well-established feature of human dialogue (Brennan & Clark, 1996), and lexical entrainment can also be measured in conversational systems (Shi et al., 2023). In deployed LLM interaction, however, apparent familiarity can enter the analysed span through several routes. D4 begins with project-file infrastructure; D3 contains public or externally formed referents; D1 uses vocabulary imported from other instances; D5 operates inside a configured companion application with persistent-memory ambiguity. Correct use of such material may be valuable, but it is not evidence that the convention formed inside the focal transcript.

This boundary changes how HCIS should be reported. Start state and source origin belong beside the mechanism scores, not in a footnote. A warm-start system can show strong mechanics from the beginning without a visible formation phase. A cold-start interaction may develop local conventions during the observed span. In a configured product, the source of an apparently long-range contribution may be impossible to recover. These details determine what the observed behaviour can support.

D3 illustrates this point. Its ordinary sampled VCS was 0.739, but fell to 0.457 when the analysis removed material formed elsewhere or assisted by public sources. This does not show that the ordinary score was wrong. It shows that “used history correctly” and “formed within this dyad” are different claims. Apparent alignment also does not prove that a pattern is unique to one partner. Constrained pseudo-dyad comparisons can test whether reuse survives a change in partner history (Wang et al., 2026). Future studies should state which claim they intend to make before scoring and use partner substitution or history removal when exclusivity matters.

7.4 Correction is an accumulation mechanism, not a courtesy marker

Immediate response to correction is commonly treated as evidence that a conversational system has recovered. Repair benchmarks appropriately test whether models can resolve a misunderstanding in the next sequence (Chiyah-Garcia et al., 2024). The focal cases show why longitudinal work must also ask what happens when the target returns.

Every case had a lower observed-window adjusted Correction Persistence value than its ACI. The size of that difference is not a calibrated capability gap because the measures use different units, opportunities, denominators, target difficulty, and censoring rules. The useful evidence lies within each case: accurate contextual interpretation, and often frequent reuse, coexisted with the recurrence of explicitly corrected behaviour. D2 provides the clearest contrast (ACI = 0.917; adjusted observed-window CPS = 0.214). D1 shows the same distinction within an otherwise strong profile.

This result gives correction a specific role in HCIS. Correction Persistence is not a politeness or responsiveness score. It records whether the interaction carries a user-defined behavioural boundary forward. Re-establishment and oscillation matter because the model may repeatedly acknowledge a correction and still repeat the behaviour. The system can appear responsive at every correction while leaving the user with a continuing repair burden.

7.5 Mechanical coherence is valence-neutral

The original manuscript associated HCIS with improved reasoning, creativity, cognitive amplification, and optimal collaboration. The present study does not measure those outcomes, and coherence is not assumed to be beneficial. Multi-turn sycophancy research shows that a model can become more conforming under sustained pressure while moving away from factual or ethical soundness (Hong et al., 2025). FIGS further separates factual integrity from calibrated validation: models may drift toward sycophancy, while factuality-only steering can instead produce cold or paternalistic responses (Pulipaka et al., 2026). HCIS mechanics and epistemic, ethical, or outcome quality therefore remain separate evaluation targets.

The qualitative examples show why successful reuse is not evidence that the reused content is valid. The model may reuse an unsupported framework correctly, and a stable convention may itself be undesirable. Outcome quality was not assessed independently. Conversely, a truthful, safe, or helpful answer can be produced without accumulated history. HCIS describes how interaction history is organised, not the moral, factual, clinical, or relational quality of the interaction.

This neutral stance prevents favourable terms such as “coherence,” “alignment,” or “engagement” from standing in for content quality. A claim that HCIS improves task performance, wellbeing, creativity, or cognitive load requires a separate outcome measure and a suitable comparison design.

7.6 Design and evaluation implications

Because the study is descriptive, the following design implications are hypotheses, not tested interventions.

First, interfaces could show where carried-forward material came from and whether it is still active. Users should be able to distinguish live context, persistent memory, retrieved notes, project files, and inferred conventions. This would reduce the ambiguity seen in D5 and help separate newly formed conventions from inherited familiarity.

Second, systems could make corrections visible and editable. A preference, boundary, or working premise could be marked as active, contested, replaced, expired, or limited to one context. The user could inspect and revise it. This would make repeated repair visible rather than leaving it as a hidden burden. Future studies must test whether such controls preserve justified corrections without freezing stale interpretations or creating rigid rules.

Third, longitudinal evaluation should report both component status and trajectory. A system that performs well on average but declines late creates a different design problem from one that starts weakly and improves. Monitoring by phase may reveal degradation that an endpoint or whole-conversation score misses. The phase rules and alert thresholds still require prospective validation.

Fourth, long-term memory evaluation could distinguish retrieval from purposeful deployment. Recall questions remain important, but natural assistance also depends on whether the model selects earlier material at the right moment and integrates it into the current task or exchange. Context Deployment offers one way to measure this. Controlled tests with and without the relevant history are still needed to show that an apparent deployment truly depends on accumulated context.

Finally, responsibility for coherent interaction should not fall entirely on the user. User behaviour, model behaviour, interface design, product memory, system instructions, and changing infrastructure all shape the outcome. The user-side construction protocol in the original paper should therefore be treated as a set of practices to test in a future intervention study, not as an established recipe.

8. Limitations and Validation Agenda

8.1 Corpus scope and generalisability

The main evidence comes from five selected interactions involving one expert user. Two human users generated the wider developmental corpus. These cases provide access to long-term conventions, corrections, reuse, and breakdown, but they are unusually invested interactions. They cannot estimate variation across users or models. The findings may depend on the focal user’s language, correction style, research awareness, interaction intensity, or selection choices. The five dyads are repeated observations, not five independent population samples. They cannot support an estimate of prevalence.

The cases also differ in product, interface, context management, memory routing, and time. These differences help expose the construct’s boundaries, but they prevent us from attributing results to a model family or architecture. Claims about typical trajectories or generality require a prospectively sampled corpus with more users.

8.2 Researcher-participant position and retrospective development

The focal participant is also the framework author and final adjudicator. Longitudinal first-person HCI can provide unusually sustained and situated access, but it does not create analytic independence and requires explicit positionality (Lucero, Desjardins, & Neustaedter, 2021). Familiarity with the logs may help identify local references. It may also increase the risk of confirmatory interpretation, attachment to particular cases, and rules refined after seeing the data.

The protocols were developed using the same natural dyads that motivated HCIS. This analysis is therefore a discovery and feasibility study, not an out-of-sample validation. Fixed sources, recorded supersession decisions, versioned protocols, and claim-level provenance reduce hidden flexibility. They do not remove the overlap between development and evaluation. Future studies should freeze protocols and hypotheses before opening held-out transcripts.

8.3 Reliability and scorer lineage

Historical coding and harmonisation used model-assisted analysis followed by checks against sources and rules. No complete independent multi-human reliability study has tested the current HCIS protocol suite. Reliability must cover which events count, where units begin and end, whether a reference depends on history, how targets are split, how phases are assigned, and which outcome label applies. Agreement only on a supplied set of rows is not enough (Artstein & Poesio, 2008).

The paper therefore does not call HCIS a “validated instrument.” A future reliability study should give independent scorers the canonical transcripts and frozen protocols. It should preserve the units each scorer detects and report agreement on both unit boundaries and labels. Human adjudication should be separate from protocol development, and model-assisted coding should disclose the scorer configuration. Source-linked evidence can make long-term judgements easier to inspect and reproduce, but repeated agreement among model judges does not prove that a label is valid (Pilditch et al., 2026).

8.4 Unequal measurement completeness

Compression uses samples of 17 to 30 User turns per case rather than every turn. Phase estimates may therefore depend on which turns were sampled, especially in short logs. Context Deployment uses different inventory types: exhaustive current-rule coding in D1 and D4, a compatible exhaustive inventory in D5, a bounded historical inventory in D3, and a conservative scripted lower bound in D2. Raw event counts cannot be compared across these inventory types. Correction Persistence depends on later opportunities and on how repeated corrections are combined into targets. Late targets may be right-censored. Phase boundaries are descriptive, not statistically inferred change points.

Functional Output Analysis is excluded because its publication rules are not frozen. Drift summarises other measures rather than adding an independent one. The tables disclose these differences, but they still limit cross-case comparison.

8.5 Source and infrastructure indeterminacy

A transcript cannot reliably separate live context, product memory, retrieval, summaries, project files, hidden system instructions, routing, or model updates. Source labels are therefore conservative and may remain indeterminate. HCIS describes behaviour at the user-facing system rather than attributing it to a base model or hidden mechanism.

Recent infrastructure studies make this limitation concrete. Persona information can remain represented in conversational history without continuing to govern first-person enactment when system-level anchoring is lost (Fraile Navarro, 2026). Memory consolidation can also preserve a proposition while changing its temporal force, making temporary information appear durable (Panthi et al., 2026). Neither result identifies the mechanism behind any focal episode; both show why retained text, faithful state, retrieved evidence, and enacted behaviour must not be treated as interchangeable.

The fresh-context question in Context Deployment is currently an analyst’s counterfactual judgement, not a controlled test. Future validation should replay matched target turns with full history, with relevant history removed, with retrieved summaries, and in clean contexts. This would test dependence on accumulated history directly and estimate false-positive deployment judgements.

8.6 Privacy, ethics, and evidence availability

The source interactions contain personal, health, creative, professional, and third-party material. Releasing them in full would conflict with participant and third-party privacy. Restricted evidence limits independent reproduction and makes the choice of quotations especially important. The final paper should use as few excerpts as possible, de-identify them, and publish only what is needed to demonstrate a decision rule. Aggregate ledgers or synthetic examples can be used when raw text cannot be shared.

Consent, independent-research status, third-party treatment, AI assistance, and provider-term review are reported in Section 4.6. No quotation enters the manuscript unless its exact source span and redaction decision are recorded.

8.7 Construct validity and unmeasured outcomes

HCIS has not yet been tested against simpler alternatives such as turn count, transcript length, repeated wording, user warmth, model verbosity, recurring topics, or application memory. The focal cases provide preliminary evidence that the components are not interchangeable. They do not establish convergent, discriminant, incremental, or predictive validity. HCIS has also not been shown to predict task success, reasoning quality, creativity, cognitive load, satisfaction, wellbeing, safety, or harm.

Validation should proceed as a programme rather than a single threshold exercise (Strauss & Smith, 2009):

  1. Independent reliability: reapply frozen protocols to held-out transcripts with independent human scorers and preserved unit detection.
  2. Fresh-context ablation: compare matched target turns under full, ablated, summarised, and retrieval-augmented histories.
  3. Partner-substitution and pseudo-dyad tests: compare focal pairings with constrained substitutions that preserve task structure while breaking partner history.
  4. Prospective correction study: introduce controlled correction targets and test immediate uptake, recurrence, generalisation, distance, topic movement, and right-censoring.
  5. Known-contrast study: prospectively sample accumulation-rich, local-fluent, warm-start, eroding, and recovery cases without selecting on the final score.
  6. Incremental validity: compare HCIS components with transcript length, lexical overlap, context position, repetition, user prompt length, and generic response-quality ratings.
  7. Cross-user replication: include diverse users, tasks, languages, products, and interaction styles; distinguish dyads from repeated observations within a user.
  8. Outcome linkage: test whether component profiles predict task or user outcomes under designs capable of supporting causal or predictive claims.
  9. Threshold research: only after reliability and validity are established, examine whether any classification rule is useful; until then, retain profiles and trajectories.

9. Conclusion

Conversation history can support accurate interpretation and purposeful reuse without producing lasting behavioural change. That is the central finding, and it explains why saying that a model “remembered” is not enough. This paper treats HCIS as a proposed family of separate, transcript-observable functions. They cover the user’s reliance on history and the model’s interpretation of it, purposeful reuse of earlier material, the persistence of corrections, and change over time. The method does not measure the quality of the current response or the truth, safety, or value of its content.

Across five purposively selected, naturally occurring dyads, the functions did not move as one. D2 retained high unpacking accuracy while most correction targets did not meet the historical observed-window persistence criterion. D1 combined strong unpacking and frequent unprompted deployment with one behavioural norm that recurred after nine corrections. Other cases showed that aggregate values concealed temporal change and that warm starts, imported material, and product memory complicated claims of within-dyad formation. These contrasts demonstrate why the component profile changes interpretation; they do not establish a validated HCIS classification or a common scale across measures.

The main contribution is therefore methodological: a traceable profile-plus-trajectory method for identifying how accumulated history shapes later human–LLM interaction and where that process breaks down. The ecological contribution is the use of interactions generated for the participants’ own purposes rather than for a benchmark. The framework still needs independent reliability testing, controlled removal of history, comparison with simpler measures, replication across users, and studies linking the profile to outcomes. The paper does more than state that conversation history matters. It offers a way to identify which longitudinal function is working, which is failing, and how the pattern changes over time.

References

Anderson, R. (2026). Sustained human–LLM collaboration as an emerging object of scientific inquiry. SSRN. https://doi.org/10.2139/ssrn.7349183

Anthropic. (2025). Consumer terms of service. Effective 8 October 2025. https://www.anthropic.com/legal/consumer-terms

Anthropic. (2026). Export your Claude data. Anthropic Privacy Center. 8 July 2026. https://privacy.claude.com/en/articles/9450526-export-your-claude-data

Artstein, R., & Poesio, M. (2008). Survey article: Inter-coder agreement for computational linguistics. Computational Linguistics, 34(4), 555–596. https://doi.org/10.1162/coli.07-034-R2

Bai, G., Liu, J., Bu, X., He, Y., Liu, J., Zhou, Z., Lin, Z., Su, W., Ge, T., Zheng, B., & Ouyang, W. (2024). MT-Bench-101: A fine-grained benchmark for evaluating large language models in multi-turn dialogues. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (pp. 7421–7454). https://doi.org/10.18653/v1/2024.acl-long.401

Bensen, W. J. (2026). Trace mutation in human–LLM dialogue: The transcript as forensic and mitigation surface. arXiv. https://doi.org/10.48550/arXiv.2604.22773

Branigan, H. P., Pickering, M. J., Pearson, J., & McLean, J. F. (2010). Linguistic alignment between people and computers. Journal of Pragmatics, 42(9), 2355–2368. https://doi.org/10.1016/j.pragma.2009.12.012

Brennan, S. E., & Clark, H. H. (1996). Conceptual pacts and lexical choice in conversation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22(6), 1482–1493. https://doi.org/10.1037/0278-7393.22.6.1482

Chiyah-Garcia, J., Suglia, A., & Eshghi, A. (2024). Repairs in a block world: A new benchmark for handling user corrections with multi-modal language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 11523–11542). https://doi.org/10.18653/v1/2024.emnlp-main.643

Clark, H. H., & Brennan, S. E. (1991). Grounding in communication. In L. B. Resnick, J. M. Levine, & S. D. Teasley (Eds.), Perspectives on socially shared cognition (pp. 127–149). American Psychological Association. https://doi.org/10.1037/10096-006

Clark, H. H., & Wilkes-Gibbs, D. (1986). Referring as a collaborative process. Cognition, 22(1), 1–39. https://doi.org/10.1016/0010-0277(86)90010-7

Deshpande, K., Sirdeshmukh, V., Mols, J. B., Jin, L., Hernandez-Cardona, E.-Y., Lee, D., Kritz, J., Primack, W. E., Yue, S., & Xing, C. (2025). MultiChallenge: A realistic multi-turn conversation evaluation benchmark challenging to frontier LLMs. In Findings of the Association for Computational Linguistics: ACL 2025 (pp. 18632–18702). https://doi.org/10.18653/v1/2025.findings-acl.958

Fraile Navarro, D. (2026). The persona is still there, but who is speaking? Latent identity reversion in persistent AI agents. arXiv. https://doi.org/10.48550/arXiv.2610.01490

Hollnagel, E., & Woods, D. D. (2005). Joint cognitive systems: Foundations of cognitive systems engineering. CRC Press. https://doi.org/10.1201/9781420038194

Hong, J., Byun, G., Kim, S., & Shu, K. (2025). Measuring sycophancy of language models in multi-turn dialogues. In Findings of the Association for Computational Linguistics: EMNLP 2025 (pp. 2239–2259). https://doi.org/10.18653/v1/2025.findings-emnlp.121

Hutchins, E. (1995). Cognition in the wild. MIT Press. https://doi.org/10.7551/mitpress/1881.001.0001

Kirk, R., Mediratta, I., Nalmpantis, C., Luketina, J., Hambro, E., Grefenstette, E., & Raileanu, R. (2024). Understanding the effects of RLHF on LLM generalisation and diversity. In International Conference on Learning Representations. https://proceedings.iclr.cc/paper_files/paper/2024/hash/5a68d05006d5b05dd9463dd9c0219db0-Abstract-Conference.html

Klowait, N., & Erofeeva, M. (2026). Nonhuman situational enmeshments: How participants build temporal infrastructures for ChatGPT. Journal of Linguistic Anthropology, 36(1), e70037. https://doi.org/10.1111/jola.70037

Kwan, W.-C., Zeng, X., Jiang, Y., Wang, Y., Li, L., Shang, L., Jiang, X., Liu, Q., & Wong, K.-F. (2024). MT-Eval: A multi-turn capabilities evaluation benchmark for large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 20153–20177). https://doi.org/10.18653/v1/2024.emnlp-main.1124

Li, J., Fu, Z., Wang, Y., Li, J., Wu, Y., & Chang, Y. (2026). What if LLMs ate their words: Causal history effects in multi-turn interaction. arXiv. https://doi.org/10.48550/arXiv.2609.05882

Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024). Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12, 157–173. https://doi.org/10.1162/tacl_a_00638

Lucero, A., Desjardins, A., & Neustaedter, C. (2021). Longitudinal first-person HCI research methods. In E. Karapanos, J. Gerken, J. Kjeldskov, & M. B. Skov (Eds.), Advances in longitudinal HCI research (pp. 79–99). Springer. https://doi.org/10.1007/978-3-030-67322-2_5

Lucero, A., Desjardins, A., Neustaedter, C., Höök, K., Hassenzahl, M., & Cecchinato, M. E. (2019). A sample of one: First-person research methods in HCI. In Companion Publication of the 2019 ACM Designing Interactive Systems Conference (pp. 385–388). https://doi.org/10.1145/3301019.3319996

Maharana, A., Lee, D.-H., Tulyakov, S., Bansal, M., Barbieri, F., & Fang, Y. (2024). Evaluating very long-term conversational memory of LLM agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (pp. 13851–13870). https://doi.org/10.18653/v1/2024.acl-long.747

OpenAI. (2026a). Exporting your ChatGPT history and data. OpenAI Help Center. Accessed 9 October 2026. https://help.openai.com/en/articles/7260999-how-do-i-export-my-chatgpt-history-and-data

OpenAI. (2026b). Terms of use. Effective 1 January 2026. https://openai.com/en-GB/policies/row-terms-of-use/

Panthi, S., Yeamin, M., Luo, S., & Abdelfattah, R. (2026). Memory consolidation flattens the temporal shape of user facts. arXiv. https://doi.org/10.48550/arXiv.2609.36457

Peng, Y., Feng, Z., Wu, T., Chen, J., Su, Y., Zhou, C., Du, T., & Ji, S. (2026). StateWise: Diagnosing and repairing persistent operational state before agent actions. arXiv. https://doi.org/10.48550/arXiv.2610.05241

Pilditch, T. D., Voudouris, K., Abbas, A., & Ududec, C. (2026). Transect: Retaining observability for long-horizon LLM agent evaluations. arXiv. https://doi.org/10.48550/arXiv.2610.08364

Portola. (2026). Terms of service. Last updated 30 June 2026. https://www.tolans.com/terms-of-service

Pulipaka, S., Binkyte, R., Sheth, I., & Abdelnabi, S. (2026). FIGS: Evaluating multi-turn sycophancy without penalizing empathy. arXiv. https://doi.org/10.48550/arXiv.2609.39863

Quan, K., Ding, Z., Yong, J., Zhang, Q., Wang, D., & Chin, J. (2026). From momentary emotion inference to sustained emotion support: Evaluating a companion agent in a longitudinal study. arXiv. https://doi.org/10.48550/arXiv.2609.16344

Shi, Z., Sen, P., & Lipani, A. (2023). Lexical entrainment for conversational systems. In Findings of the Association for Computational Linguistics: EMNLP 2023 (pp. 278–293). https://doi.org/10.18653/v1/2023.findings-emnlp.22

Sonar, D. (2026). LSREP: A longitudinal state-replay protocol for evaluating conversational memory, with ICE v2 as an audited local-first architecture. arXiv. https://doi.org/10.48550/arXiv.2609.16730

SpaceXAI LLC. (2026). Terms of service: Consumer. Last updated 11 September 2026. https://x.ai/legal/terms-of-service

Strauss, M. E., & Smith, G. T. (2009). Construct validity: Advances in theory and methodology. Annual Review of Clinical Psychology, 5, 1–25. https://doi.org/10.1146/annurev.clinpsy.032408.153639

Testeza, T., & Solomon Research Group. (2026). Fork, recovery, and historically organized continuity in a longitudinal LLM–human interaction: A naturalistic case study. Zenodo. https://doi.org/10.5281/zenodo.22803892

Wang, P.-Y. A., Mishra, C., Özyürek, A., Rubio-Fernández, P., & Ghaleb, E. (2026). Aligned but not partner-specific: How multimodal LLM agents succeed in reference games without forming conceptual pacts. arXiv. https://doi.org/10.48550/arXiv.2606.08081

Wojewodzka, A. (2026). High-Coherence Interaction State (Third Space) as An Interactional Regime in Human–LLM Dyads (Version 1.1). Zenodo. https://doi.org/10.5281/zenodo.18432386

Yu, H., Koh, H., Kim, M., Jang, Y., & Jung, K. (2026). PRAGMA: Evaluating personalized guidance with memory alignment in lifelong conversations. arXiv (accepted at EMNLP 2026). https://doi.org/10.48550/arXiv.2609.09664