Thinking of ACE? We Can Do It with Fewer Tokens

Details

IBM Research's blog post on Hugging Face introduces ALTK-Evolve, a system that lets AI agents accumulate operational knowledge from their own past task trajectories without weight updates or human labels. The comparison target is ACE (Agentic Context Engineering), a recently published system that takes the same general approach. Both systems share a core principle: don't compress what an agent learns. ACE uses per-bullet counters on an evolving playbook; ALTK-Evolve maintains individually retrievable guidelines with support counts. In both cases, a lesson that emerged from five independent episodes is treated as fundamentally different from one that appeared once — compression destroys that signal. Where they diverge is delivery. ACE prepends its full playbook to every inference call — richer context at the cost of token consumption that scales with playbook size. ALTK-Evolve retrieves only the guidelines relevant to the current task, keeping the context window lean. IBM reports comparable task performance at lower token cost, which matters most in long-running agent loops where each step incurs inference cost. The target domain is enterprise multi-app tasks: splitting bills, reconciling orders across nine simulated tools, finding songs through API chains. Realistic agentic scenarios where accumulated operational knowledge meaningfully improves completion rates without touching model weights.