Persistent State Helps; Function-Specific Ownership Shows No Material Advantage Against Relevance-Matched Shared Persistence

Systems that give language models persistent state often keep it private to the agent, role or function that produced it. This study tests whether such ownership helps beyond relevance. In synthetic hidden-rule induction tasks, one open-weights model answered questions with ten persistent records in context. The records were produced outside the model by scripted executors under different task functions' rules. Four conditions were compared: OWNED: the evaluated function's own records; SHARED: records drawn origin-blind from all functions and matched to OWNED's relevance profile, record count, sufficiency and context length; CROSS: the same profile drawn from other functions only; RESET: format-matched records from an unrelated filler world, certified to carry no information about the task. The analysis, result classes and a materiality margin of ±.099 were fixed before execution. 2,934 frozen calls ran in two blocks with no interim inspection, and an independent reviewer recomputed the analysis from the raw records. Persistence helped substantially: each persistent condition beat RESET by about .32–.37 in success rate. Ownership showed no material advantage: OWNED − SHARED was +.027 (95% CI −.042 to .096; p = .49; 232 units in 203 worlds), so the prespecified class is "persistence effect without an ownership advantage". Advantages smaller than about .096 are not excluded, and the result does not concern memory that a model writes for itself. Licence scope. CC BY 4.0 covers the manuscript (PDF and Markdown source), the README, the results table and the per-unit outcome table. recompute.py is licensed MIT. NOT released: the experiment software, prompts, task vocabulary and records, raw model responses, and internal design and review documents. AI-assistance disclosure. The author used AI systems, including ChatGPT and Claude, as research-assistance tools for evidence organisation, source comparison, analysis support and manuscript drafting. The experiment apparatus, analysis code, independent reviews and this manuscript's drafting were carried out by AI agents under the author's written directions. Experimental design decisions, governance rulings, interpretation of results, selection of claims and final responsibility remain with the author. Every number was checked against the preserved study records.