Large language model (LLM) agents excel in interactive settings, yet their capacity for continual evolution from experience is limited. Fine‑tuning can adapt agents but requires access to model parameters and incurs high computational cost, which is impractical for large‑scale or closed‑source LLMs. External memory allows agents to store experience without altering parameters, but existing approaches focus on representation and organization, leaving the acquired knowledge tightly bound to specific tasks and contexts and thus poorly generalizable. The key challenge is converting concrete interactions into abstract, reusable knowledge that can guide future decisions beyond individual episodes.
We introduce SAGA (Self‑evolving Agents through Experience‑Grounded Abstraction), a framework that abstracts experience into hierarchical knowledge and leverages it during inference. SAGA progressively transforms interaction trajectories into episodic narratives, reusable procedures, and principles equipped with explicit applicability conditions, while preserving links to execution evidence. Retrieved principles are instantiated in the current task context to produce concrete guidance, which refines candidate actions through corrective feedback and resampling. This creates an execution‑abstraction feedback loop: accumulated knowledge steers future interactions, and new experiences continuously update the hierarchical memory.
Experiments on ScienceWorld and ALFWorld show that SAGA improves task performance markedly. Ablation studies highlight that contextual instantiation of principles and action regulation are essential for exploiting principle‑level knowledge.
Review: SAGA provides a viable route to abstract experience into reusable principles via hierarchical memory, offering a promising direction for continual learning of large models in closed‑source settings.