Thinking of ACE? We Can Do It with Fewer Tokens

Hugging Face has published a comparison between its ALTK-Evolve system and ACE, two agentic memory frameworks that allow LLM agents to learn from their own trajectories at inference time without requiring weight updates or human labels. Both systems avoid summarizing away hard-won lessons, instead utilizing counters to track experience and support for individual guidelines. However, the two frameworks differ significantly in their delivery mechanisms. ACE injects a comprehensive, evolving playbook into the model on every single step regardless of context. In contrast, ALTK-Evolve treats delivery as a dial, providing a small core of guidelines extended per task or serving the full consolidated set depending on model headroom and capability. According to Hugging Face's in-house evaluations on the AppWorld benchmark using DeepSeek-V3.2 and gpt-oss-120b base models, this delivery difference drastically impacts the token bill. ALTK-Evolve achieved matching or superior task goal completion accuracy at roughly 40 percent of ACE's inference cost on the strong model, and at about one-seventh the cost on the weak model. Hugging Face attributes these savings to selective retrieval, which prevents weaker models from being overwhelmed by large contexts while preserving accuracy on difficult tasks.