ALTK-Evolve vs ACE: Calibrated Agentic Memory Cuts Inference Cost

IBM Research introduces ALTK-Evolve, an agentic memory system that improves LLM agent performance on multi-step tasks by learning from past trajectories without weight updates or human labels. The system is compared to ACE (Agentic Context Engineering), with both agreeing that compressing an agent's lessons into a tidy summary is counterproductive. Instead, they count lessons rather than collapse them. The key difference lies in delivery: ACE injects a comprehensive playbook at every step, while ALTK-Evolve calibrates delivery based on the model's capacity, sending a small fixed core of high-support guidelines plus a task-specific selection, or the full set when the model can handle it. On the AppWorld benchmark, ALTK-Evolve achieves same-or-better accuracy at roughly 40% of ACE's inference cost on a strong model (DeepSeek-V3.2) and about one-seventh the cost on a weaker model (gpt-oss-120b), with a tie in accuracy (56.0 vs 54.8). The by-difficulty breakdown shows that selective retrieval wins on hard tasks, while full playbooks help on easy and medium tasks for the weaker model. The system uses support counts, typed guidelines (strategy, recovery, optimization), and causal attribution with provenance. The library and technical report are available. The comparison is controlled by running ACE in-house on the same base models and splits.