CMU AI Agents 2026: 4. Memory and Skills for Agents

Agents improve performance on recurring tasks by storing and reusing knowledge outside their immediate context window through skill induction. This process involves transitioning from raw action sequences to structured, reusable representations like text-based workflows or programmatic functions. While human-authored skills provide high-quality, interpretable guidance, automated skill induction allows agents to abstract best practices from past successful trajectories. Effective management requires balancing memory size with relevance, often utilizing progressive disclosure or retrieval mechanisms to avoid context window saturation. Research in this field, including frameworks like MemGPT and SkillsBench, highlights that while code-based skills offer modularity and testability, they require robust error handling to remain generalizable across diverse environments. Ultimately, integrating these memory systems enables agents to solve increasingly complex, multi-step tasks by leveraging learned subtask abstractions rather than relying solely on individual, low-level actions.