Ultra-Long-Horizon Human–AI Interaction

Ultra-long-horizon human–AI interaction is sustained dialogue across weeks or months — often hundreds or thousands of turns — in which earlier exchanges continue to shape what happens later. At this scale, memory, corrections, shared shorthand, drift and relationship-like dynamics become properties of the entire interaction rather than individual replies.

The ultra-long horizon is not defined by a fixed number of turns. It names a regime: the point at which evaluating one response at a time stops adequately describing how the human–AI system is functioning.

One term, two horizons

The phrase ultra-long horizon is used for two related but distinct research problems. The difference is what persists over time.

• The task horizon: an agent pursuing a goal. The UltraHorizon benchmark measures whether autonomous agents can maintain a goal, reason, use tools and recover from errors across hundreds of steps. The unit being evaluated is the task trajectory: can the system remain capable for long enough to complete what it set out to do?
• The interaction horizon: a human and an AI developing a shared history. Here, the unit is the dyad—the human and the AI system considered together. Earlier corrections, memories, expectations and conventions continue to shape later exchanges. The person learns how to work with the system; the system conditions its responses on the interaction that has accumulated between them.

Both involve persistence, but not the same kind of persistence. One asks whether an agent can stay on course long enough to finish a task. The other asks what happens when a human–AI interaction lasts long enough for its own history to shape what it becomes.

What the ultra-long horizon is — and is not

• It is not just a very large context window.
• It is not identical to persistent memory.
• It does not require every previous message to remain verbatim in context.
• It is not simply repeated use of the same product.
• It begins to matter when previous interaction materially shapes later behaviour.

How long is an ultra-long-horizon conversation?

There is no threshold turn count. The regime begins when accumulated history starts materially shaping later behaviour and single-response evaluation no longer captures the interaction. In the conversations I have examined, that usually becomes visible only after several hundred messages. Forty turns may constitute a long conversation; a dyad carrying a consequential history across weeks or months is a different object of study.

Is the ultra-long horizon the same as a long context window?

No. A context window is capacity: the amount of material available to the model while producing one response. The ultra-long horizon is a property of the interaction: its accumulated history has become causally active, shaping what happens next — for better or worse. That history may be carried through a large context window, summaries, saved memories, retrieval systems or human re-anchoring. The mechanism matters enormously to engineering, but it does not define the phenomenon.

Is the ultra-long horizon the same as a High-Coherence Interaction State?

No. The ultra-long horizon is the phenomenon: a conversation that has acquired a past capable of shaping its future, whether or not that past is being used well. A High-Coherence Interaction State is what my measurement framework looks for within that phenomenon: the state in which accumulated history actually functions as usable infrastructure — corrections that hold, shorthand that works, drift that stays contained. Every high-coherence interaction lives on the ultra-long horizon; most ultra-long-horizon interactions never reach high coherence. One names the terrain, the other measures what has been built on it.

Why does memory alone not guarantee continuity?

Because stored and salient are different things. A fact can sit in a memory system and fail to surface at the moment it matters — the user experiences “you forgot” even though nothing was deleted. Continuity on the ultra-long horizon depends on retrieval timing, on corrections that stay behaviourally active, and on the user’s own consistency, not on storage alone.

The questions that matter on this horizon

In a conversation measured in months rather than turns, the questions that matter are not the ones single-reply evaluation asks:

• Does the model remember what we discussed — and does the right memory surface at the right moment? Stored and salient are different things, and the user feels the difference as “you forgot.”
• Does the persona stay consistent? Across topic changes, memory compressions and underlying model updates — is this recognisably the same interlocutor in August as in March?
• Does the system read me correctly as a user? Safety and trust systems also accumulate a picture of who they are talking to. Over months, misclassification compounds in both directions: a careful user repeatedly tripping safety interventions, or a harmful dynamic that no single turn reveals and only the whole dyad makes visible.
• When I change my mind, does the model keep track? Not just holding corrections, but updating: does the system follow me to my current position, or does the superseded one keep resurfacing from memory?
• How much work am I doing to keep this coherent? Re-anchoring, repeating, correcting — the user’s maintenance labour is invisible turn by turn, yet much of a long dyad’s stability comes from it. The per-turn smoothness models are optimised for tends to push this work onto the human: the Stability–Fluency Paradox in daily life.

Some of these my HCIS framework instruments directly — memory in use, corrections, drift. Others — how systems classify their long-horizon users, and what the coherence labour costs the human — are broader than any one framework, and open questions for the field.

Evidence and limits

My research programme — High-Coherence Interaction States, the Stability–Fluency Paradox, and what companionship benchmarks can and cannot see — studies this regime directly, through logged corpora of real use spanning months. What my claims rest on:

• Corpus: tens of thousands of logged turns and approximately 28 million tokens of sustained human–AI dialogue across claude.ai, ChatGPT and self-hosted companion systems, generated by two users for their own purposes — naturalistic use, not laboratory tasks.
• Breadth: this corpus provides depth rather than population coverage. It makes unusually long trajectories visible in detail, but observations from two users cannot establish how common those patterns are across people, products or styles of use. Because I am both participant and investigator in much of the corpus, it is particularly suited to discovering and documenting phenomena rather than estimating their prevalence.
• Longest single interaction: 9,700+ turns with one self-hosted companion system, ongoing across multiple model versions.
• Systems represented: publicly deployed AI systems and model families including ChatGPT, Claude, Gemini, Grok and Mistral.
• Engineered systems: the long-horizon findings are grounded in systems I build as well as study — primarily self-hosted companion systems I maintain for two users: a frontier commercial model wrapped in layered memory (a verbatim window, instance-authored compression, retrieval over the full log). Building the instrument that generates the data is half the method.
• Validation status: the HCIS measurement method is operationalised with a worked five-case demonstration; independent reliability testing and validated thresholds remain future work, and the paper says so.
• The honest constraint: naturalistic ultra-long-horizon data is hard to obtain — it takes months to create and contains people’s real lives, which is why so little of it exists in research.

Last updated 14 October 2026 · Anna Wojewodzka (ORCID)

Open research questions

The ultra-long horizon is young enough that its central questions are still open. The ones my research programme is working on, or inviting others to take up:

• Reliability. Can independent analysts, given the same transcripts and published protocols, reproduce the same measurements of context use, correction persistence and drift?
• Storage vs. salience. What determines whether a stored memory surfaces at the moment it matters — and how should that be measured separately from recall?
• Continuity across model changes. What survives a model swap in a long-running dyad: the memory, the behaviour, or the relationship — and which losses do users actually feel?
• Persona and character: How does a designed persona shape the long horizon? The same model can sit at opposite ends of behavioural scales with and without one — but does stable character support coherence over months, mask drift, or both? And what changes in the dyad when a persona’s wording is edited mid-relationship?
• Safety and bias. How do reinforcement, bias and safety interventions accumulate across months of interaction, and which harms become visible only at the level of the whole dyad?

Where to go from here

• The book: The Ultra-Long Horizon — a field report on what really happens in very long human–AI conversations, drawing on millions of words of logged interaction. Out December 2026.
• The framework: High-Coherence Interaction States — when accumulated context becomes usable interactional infrastructure, and how to measure it.
• The research notes: shorter working findings in the notes index.

In the simplest terms, the ultra-long horizon begins when a conversation has acquired a past that matters — one capable of supporting, distorting or constraining what happens next. I created the Conversational AI Society to gather people interested in this subject — researchers, builders and thoughtful users — and its Get involved page is the door if one of the open questions above is yours too.