Hugging Face AI Updates: August 12, 2026

1. IBM Research Releases ALTK-Evolve, a Token-Efficient Alternative to ACE for Agent Memory

Hugging Face. IBM Research published ALTK-Evolve on the Hugging Face blog, an open-source library that teaches LLM agents from their own experience by extracting, consolidating, and retrieving reusable guidelines from agent trajectories without human labels or weight updates. Unlike the ACE framework, which injects a full playbook at every inference step, ALTK-Evolve selectively retrieves only the guidelines relevant to the current task. On the AppWorld benchmark it reported higher accuracy than ACE while using roughly 40% of the tokens, and matched ACE’s accuracy on weaker models at about one-seventh the inference cost. Source