2026 Survey Offers Unified Framework for Understanding LLM Agent Memory

A 2026 survey by Tang et al. proposes a unified framework for understanding memory in large language model agents, addressing the fragmented landscape of existing approaches such as vector databases, retrieval pipelines, and continual learning. The framework organizes memory into three representation types — token-based, intermediate, and parameter memory — treating them as complementary layers rather than competing methods. It identifies three core memory operations: constructing what is worth storing, updating memories as new information arrives, and querying the right memory when needed. The survey also emphasizes that forgetting is a necessary feature, with pruning, summarization, and expiration policies keeping memory systems efficient and accurate. The central argument is that AI memory is shifting from passive storage to an active lifecycle, where agents continuously decide what to remember, update, retrieve, and discard.