Tsinghua and Bosch Publish an Architecture-Centric Taxonomy of LLM Memory Along Three Axes: Representation, Update Dynamics, Persistence

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

Sining Zhoubian, Dan Zhang, Evgeny Kharlamov, and Jie Tang's July 28 survey organizes the sprawl of LLM memory work by representation (implicit versus explicit), update dynamics (offline versus online), and persistence (short-term versus long-term), then formalizes the mechanisms every system implements ad hoc: memory writing, routing, state transitions, and consolidation. At 20 pages with 4 figures it is short enough to actually read, and the named open problems — hybrid memory architectures, system-level efficiency trade-offs, and multi-dimensional evaluation methodology — line up with where practitioners keep improvising. Useful as a design checklist if you are choosing between vector store, knowledge graph, and in-weights approaches for an agent's memory rather than accreting all three.