Caring for the system that cares for me: An autoethnography of designing sustained memory with a stateless conversational AI
Ziyuan Jiang · Design and Artificial Intelligence · 2026
Large language models are stateless by design. Yet users form lasting relationships with them, and when a platform does not provide persistent memory, the work of sustaining continuity can fall to the user. This paper is a six-month analytic autoethnography of that work. Between July 2025 and January 2026, I built and lived with River : a memory relay system developed and maintained across eighteen iterations of a stateless conversational AI, implemented as a Claude Project containing markdown files for core rules, a style guide, appendices indexed for on-demand retrieval, and relay documents that handed conversational state from one session to the next. I trace the trajectory across five phases and identify four cross-cutting dimensions of relational work: translation, curation, calibration, and negotiation. Three structural paradoxes recur across the trajectory. Self-archiving , in which curating what will be remembered trades away memory’s capacity to surprise. Caregiving reversal , in which the system built to support its user requires continuous user support to remain coherent. Gardening , in which the relational qualities that matter most can be cultivated but never engineered. Alongside these, the paper develops a distinction between profile-based and relay-based continuity, between being known about and being known where , and argues that the dominance of the first in current memory research and commercial deployment has obscured the second. I situate these findings at a specific historical moment in commercial AI memory deployment, and argue that autoethnography is an essential complement to engineering-oriented accounts of conversational AI memory.