Retained memory and fresh exploration under a matched evaluation budget in a changing-target population model

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Memory can improve search relative to one alternative while losing to another at the same evaluation budget. We examine this distinction in an abstract changing-target model with 32-bit states, 32 individuals and 40 updates. Every policy makes 64 charged objective queries per update. A shared four-record historical cache is compared with current-population donors, an allocation that replaces repeated current probes with children, and four fresh uniform scouts. Historical retrieval with current-parent fallback improves cumulative accuracy over a current-child comparator under independent targets, yet loses to uniform scouting. Alternation between two targets reverses the memory-versus-scout ranking in the first cohort. A separate cohort supports the recurrence interaction and the independent-target disadvantage, but leaves recurrent cumulative superiority unresolved. Thus the full ranking-switch prediction is only partially replicated. On the second, already observed cohort, one fixed memory/scout hybrid improves cumulative accuracy against both named regime references; its equal-weight mean gain is 0.925 percentage points (approximate pointwise 95% interval, 0.621 to 1.210). With scouting and candidate allocation held fixed, replacing current donors by retained history increases cumulative accuracy by 2.158 points under recurrence and 0.159 under independent targets. The recurrence-minus-independent interaction is 1.999 points (1.412 to 2.617). Terminal comparisons and individual failures qualify these means. The findings support conditional search value of an imposed retention policy. They do not establish a universally superior optimizer, selection for memory, spontaneous memory origin or a biological effect. Research preprint, version 1.0.0; not externally peer reviewed. This deposit includes the manuscript, supporting information and a compact evidence package containing all nine saved statistical grids, block summaries, bootstrap rows, archived scientific source, figures and reproducibility scripts. Large raw trajectories and random tapes are excluded. No new population simulations were run for publication. Codex assistance is disclosed. Manuscript, figures and data: CC BY 4.0. Original code: MIT. Versioned repository release: https://github.com/jackchenx3/memory-and-fresh-search/releases/tag/v1.0.0 This shared-cache study has a distinct evidence scope from the private-memory strategy-transmission preprint, https://doi.org/10.5281/zenodo.22907237, and the two-dimensional population-state preprint, https://doi.org/10.5281/zenodo.22907602.