This survey provides an architecture-centric taxonomy of LLM memory across three axes: representation (implicit vs. explicit), update dynamics (offline vs. online), and persistence (short-term vs. long-term). It categorizes advances in attention, recurrent states, and scalable lookup storage.
HOW THIS AFFECTS YOU
●
builderYou can use this taxonomy to better select memory implementation strategies like RAG or recurrent states for your applications.
●
researcherThis framework helps organize the fragmented landscape of memory-related architectural research.