Event

Proposes MCCL Memory-Centered Continual Learning Framework

1 reports · 1 sources

A paper proposes a memory-centered continual learning (MCCL) framework, arguing that the quality of memory should be judged by whether it improves an agent's subsequent actions, rather than only by what facts are retrieved. The framework consists of four parts: a three-axis taxonomy organizing memory by storage location, form, and management approach; a dual-loop lifecycle in which an inner loop updates memory states and an outer loop improves management strategies based on task feedback; an evaluation protocol measuring memory retention under interference, adaptation to change, and task transfer, used to distinguish genuine behavioral gains from mere context expansion; and a memory scaling framework characterizing how agent performance scales with experience, storage capacity, and compute budget across non-parametric and parametric memory.