Thinking of ACE? We Can Do It with Fewer Tokens

Source: HuggingFace Date: 2026-08-11 URL: https://huggingface.co/blog/ibm-research/altk-evolve-sldd

Summary

IBM Research’s ALTK-Evolve replaces ACE’s (Agentic Context Engineering) practice of injecting a full evolving playbook at every agent step with selective, clustered, causally-attributed retrieval of just the relevant guidelines. On AppWorld benchmarks it matches or beats ACE’s task-completion rate at 15-40% of the token cost (89.3% vs. 80.4% at 263K vs. 634K tokens on DeepSeek-V3.2; 56.0% vs. 54.8% at 116K vs. 777K tokens on gpt-oss-120b).

Implications

Feeds the agent-layer-convergence thread — a context/memory-engineering technique competing directly with the ACE pattern already circulating in agent frameworks. Token-efficiency gains at this magnitude matter for anyone running persistent agent memory in production; worth tracking whether orchestration frameworks adopt selective retrieval as a default over full-playbook injection.