ALTK‑Evolve: On‑the‑Job Learning for AI Agents

IBM Research has unveiled ALTK-Evolve, a novel long-term memory subsystem designed to solve a critical flaw in current AI agents: their inability to learn from experience. Most agents operate like an 'eternal intern,' re-reading transcripts of past interactions without distilling reusable principles, leading them to repeat mistakes and fail to adapt. ALTK-Evolve addresses this by implementing a continuous loop that observes agent trajectories, extracts candidate guidelines, refines them through scoring and consolidation, and injects only the most relevant ones back into the agent's context for future tasks. This transforms one-off events into portable strategies.

In rigorous testing on the AppWorld benchmark—where agents perform realistic multi-step tasks via APIs—the system demonstrated significant gains. A ReAct agent equipped with ALTK-Evolve showed an 8.9% aggregate improvement in Scenario Goal Completion (SGC), a strict consistency metric. The most dramatic improvement was on hard tasks, where success rates jumped by 14.2 percentage points, a 74% relative increase. The system proves that agents can generalize learned principles to unseen tasks, with performance gains scaling with task complexity. By filtering noise and providing just-in-time guidance, ALTK-Evolve moves AI agents beyond simple prompt-following toward accumulating actionable, environmental wisdom.