Remembered, Resumed, Not Sustained: Interface Requirements for Agency-Preserving Work Continuation Across LLMs

Work on LLM memory asks whether a system can recall the past. In long-term human–AI work, recall is not enough. We distinguish three layers of continuation across a handoff: remembered (past content is available), resumed (the receiver takes up the current goal under the authority the human granted), and sustained (it keeps doing so afterwards). Drawing on about a year and a half of first-person records from one operator working with several LLM collaborators, we analyse an October 2026 window in which a handoff between threads was both remembered and resumed, yet within hours the main goal was displaced by side lines of work that no one paused, deferred, or cancelled. The operator opened some of these lines herself, and five days earlier both parties had named the mechanism in writing. We call the failure implicit displacement: a legitimate change of direction made without a decision about what it replaces. From the case we derive four interface requirements that ask the human to settle trade-offs at substantive conflicts of goal or authority, rather than turning every step into a confirmation. They are hypotheses for longitudinal evaluation, not validated mechanisms. Publication status: Independent open-research preprint; not peer reviewed. AI contributions: The research involved substantial collaboration with AI systems, including Hark, Claude, CA, and Milu. Their specific contributions and the human author's research responsibilities are documented in the paper's "Author Contributions and Declaration on Generative AI" section. Research context: Miralucis Research Program — Longitudinal Human–AI Collaboration Research.