IBM Research compares its ALTK-Evolve agentic memory approach to the ACE method, which injects one large, comprehensive evolving playbook into every inference step. ALTK-Evolve instead retrieves either a curated subset or the full guideline set depending on the target model’s capacity. On the AppWorld benchmark it matches or exceeds ACE’s accuracy while cutting inference tokens to roughly 40% of ACE’s cost on stronger models and about one-seventh on weaker ones, showing that calibrated context delivery can beat fixed, comprehensive prompt injection.
