ALTK-Evolve Cuts Agent-Memory Inference Tokens Against ACE

Summary

IBM Research compares ALTK-Evolve and ACE, two systems that learn reusable lessons from LLM-agent trajectories without changing model weights. ACE sends its full playbook at every step, while ALTK-Evolve retrieves task-relevant guidelines and adjusts context size to model capability. In AppWorld tests, ALTK-Evolve achieved 89.3% task goal completion on DeepSeek-V3.2 using 263K tokens per task, versus ACE’s 80.4% with 634K tokens. On gpt-oss-120b, it scored 56.0% versus 54.8% while using 116K instead of 777K tokens.