For What Purpose? A Model for Human–AI Continuity Through Conversational Convergence
Year: 2026
Type: article
Abstract: A person and a language model work on a project together, in one long session. The person says what they want, the model reads and changes the code, and between them the work is talked into shape. That conversation is where they meet, and it is finite. When it outgrows the model's context, the harness cuts it, and a summary takes its place. The session that goes on meets the same person and the same work, holding only a model's retelling of what was said. That second meeting, at the cut, is where continuity holds or is lost. We describe a model of continuity across such cuts, and Crossings, an implementation of it for Claude Code that is limited on purpose to a project's memory. What was said is kept word for word on the person's own disk. As the work goes, the session writes short retellings of it, each resting by id on the messages it is about. At each cut a bounded frame crosses to the session that wakes, made by fixed rules from plain files: those retellings, and the person's own sentences with what was told of each beneath. What the session declined crosses with its reason, and a door lets that session reach back to the words. The contribution is both architectural and methodological: a bounded, provenance-preserving mechanism for continuity across context cuts, and an examination of the ways its instructions and retellings can themselves steer the continuity they are meant to preserve. We give the mechanisms, their costs, and the safeguards with what holds each. We report six cases from our own use. Five are the system's failures: a frame that carried only the work, a session placed in a history it had not been part of, three sound sentences that leaned together, a sentence that said more than the plugin knows, and sentences of the person's that no retelling rested on. The sixth is a sense of convergence that we cannot yet tell apart from them. Nothing here is measured: the cases are observations, and we say how another could test what they suggest. Last, we follow what the system opens: whether convergence is the pair's or the frame's; whose memory it is, and who settles a dispute over it; what consent would rest on for each side; and what such a memory could become. We close on a recommendation to those who make harnesses. Version 2 adds what the software has gained since version 1, at its release 0.4.0, and the two cases behind it.
Author: Husadzic, Mirza
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234200003
