New ALMA Framework Uses Meta-Learning to Auto-Design Memory for AI Agents, Outperforming Hand-Built Systems

Researchers have introduced ALMA, a framework in which a meta-level agent writes memory designs as executable code and discovers architectures that consistently beat human-engineered ones across four decision-making benchmarks. The work, titled "Learning to Continually Learn via Meta-learning Agentic Memory Designs," was published on arXiv by Yiming Xiong, Shengran Hu, and Jeff Clune, and its code has been released as open source.