Dispatch
IBM Research ALTK-Evolve: On-the-Job Learning System That Turns Agent Trajectories into Reusable Guidelines
IBM Research published ALTK-Evolve on Hugging Face, a system that addresses the 'learning gap' in AI agents by building long-term episodic memory from execution traces. The system captures full agent trajectories, mines them for structural patterns via pluggable extractors, then consolidates, scores, and prunes to generate reusable guidelines. In benchmarks, ALTK-Evolve improved reliability 14.2% on AppWorld's hard tasks without context bloat, and integrates with OpenAI, LiteLLM, and HF agents.
↳ Follow the thread