Prospective ablation study reveals persistent memory wikis reduce context and latency without improving task quality in AI agents, highlighting efficiency gains over accuracy.
A prospective three-condition study of no Wiki, flat history, and persistent Wiki for governed Agent Skill evolution. The release includes the paper, frozen-result summaries, a verified 36-entry Skill Impact Ledger, and reproducibility checks for 87 completed model calls. Persistent Wiki reduced context, latency, and final rule length relative to flat history, but did not improve mean task quality in this narrow study. Source code and machine-readable protocols are Apache-2.0; the manuscript, documentation, result tables, figures, and release metadata are CC BY 4.0.
No takes yet. Share an insight, caveat, or question.
Song Luo (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: