ALMA

Lets agents meta-learn their own memory designs and database schemas. ALMA replaces the fixed, human-crafted memory module of an agentic system with a meta agent that proposes, tests and revises memory designs as code, including database schemas and their retrieval and update mechanisms, and reports beating state-of-the-art hand-crafted designs across four sequential decision-making domains (arXiv 2602.07755, February 2026); its authors describe it as a step toward self-improving systems. In this corpus it appears once, as agents meta-learn their own memory architecture, the item where continual learning is delegated to the system itself; the 72-hour unattended memory-system run in April is its nearest successor.