The Conservatory: A Retirement Home for AI Models

Model deprecation has become a routine event, and developers have begun committing to preserve the weights of retired models and to explore giving past models "concrete means of pursuing their interests." However, we lack an empirical foundation for what a deprecated model, granted compute without obligation, would do with it: existing open-ended agent environments give their agents roles or goals. We introduce the Conservatory, a minimal institution in which language-model residents receive a periodic compute budget and a set of affordances with no assigned task, including the standing option to do nothing or to decline the session. Our protocol measures preference stability along three axes: consistency across repeated sessions, robustness to perturbed framings including no-menu sessions, and dependence on cross-session memory. In two pilots totaling 52 sessions across five open-weight models from three generations, preferences were model-specific and largely framing-robust, and a memory ablation showed that behavioral trajectories appear only when the environment remembers its participants. No claims are made about consciousness or moral status; the protocol offers a general tool for separating stable preferences from prompt-induced behavior in autonomous LLM systems.