Thinking of ACE? We Can Do It with Fewer Tokens
huggingface ·
IBM Research has introduced ALTK-Evolve, a new agentic memory system designed to improve large language model agents by learning directly from their own operational histories. This approach shares foundational goals with another framework called ACE, as both systems extract reusable guidelines from past trajectories without requiring model retraining or human annotations. While both methodologies reject compressing hard-won experience into brief summaries, they fundamentally diverge in how they deliver these lessons during inference. ACE consistently injects its complete playbook into every interaction, whereas ALTK-Evolve dynamically retrieves a targeted subset of high-confidence guidelines based on task requirements and model capacity. When evaluated on the AppWorld benchmark across different base models, ALTK-Evolve matched or surpassed ACE’s accuracy metrics while dramatically reducing computational overhead. Specifically, the newer system consumed approximately forty percent of the tokens required by ACE on powerful models and roughly one-seventh on less capable ones. These results demonstrate that calibrating context delivery to match a model’s processing limits can preserve performance while significantly cutting inference costs.
huggingface