ALTK-Evolve, a system for agentic memory that converts agent trajectories into reusable guidelines without weight updates, matches or exceeds the accuracy of ACE while using significantly fewer tokens. Unlike ACE, which injects a comprehensive playbook on every step, ALTK-Evolve uses selective retrieval to deliver only relevant guidelines based on model capacity.

  • On DeepSeek-V3.2, ALTK-Evolve outperforms ACE on Easy, Hard, and Overall tasks at approximately 40% of the inference cost.
  • On gpt-oss-120b, it ties ACE with a score of 56.0 to 54.8 at about one-seventh the cost.
  • The system avoids compressing lessons into summaries, instead keeping distinct guidelines with support counts to preserve experience.
  • Delivery is calibrated: strong models receive the full consolidated set, while weaker models receive a curated subset to avoid context crowding.

By treating memory delivery as a dial rather than a constant, ALTK-Evolve prevents guidance from overwhelming weaker models and reduces token usage on stronger ones.