Tsinghua, Bosch Researchers Map LLM Memory Onto Three Axes

Researchers at Tsinghua and Bosch sort large language model memory into three axes, representation, update dynamics, and persistence, in a new arXiv survey.

That's the taxonomy most memory systems skip. Teams build whatever mix of vector store, knowledge graph, and fine-tuned weights feels right, then patch the mismatches once things break in production.

The paper formalizes four mechanisms most systems build ad hoc: memory writing, routing, state transitions between working and long-term storage, and consolidation. It runs twenty pages with four figures, short enough to read straight through instead of hunting for one diagram.

Its open-problems list, hybrid architectures, system-level efficiency tradeoffs, multi-dimensional evaluation, reads like a checklist for picking a memory design before writing code.

I didn't have this when I built mine. I stacked a vector store, a knowledge graph, and fine-tuning together, and I'm still working out which one owns which fact. Pick a representation before you start stacking. Sorting out the overlap after the fact costs more than the paper does to read.