A Three-Axis Taxonomy for Memory Architectures in Large Language Models
An arXiv survey, dated 29 July 2026, proposes a systematic, architecture-centric taxonomy of memory in large language models.
Shane Barrett·updated July 30, 2026
The framework characterizes memory along three orthogonal axes—representation, update dynamics, and persistence—and maps the trajectory of memory-centric LLM design against system-level efficiency trade-offs.
Three-Axis Taxonomy
The taxonomy decomposes LLM memory into three independent dimensions. Representation addresses how information is encoded within a model's parameters and activations. Update dynamics governs how stored state is modified during training and inference. Persistence captures how long information remains available—across a single context window, across sessions, or across the model's operational lifetime. Treating the axes as orthogonal positions existing approaches on a shared coordinate system rather than evaluating each in isolation.
System-Level Trade-offs
Memory-centric LLM design carries direct efficiency costs. Choices along the persistence axis create tension between retrieval latency, parameter overhead, and effective usable context. A separate explainer published the same day in Business Day reinforces the practical stakes: a 200-page set of financial statements corresponds to roughly 150,000 tokens, and languages other than English consume several times more tokens per equivalent meaning, accelerating context-window saturation and raising inference cost. Memory mechanisms that extend effective context inherit these asymmetries. Designing around bounded capacity—whether in compute budgets or in training periodisation—forces explicit choices about when state is encoded, updated, and allowed to fade.
What Remains Unverified
The arXiv entry exposes only the abstract framing. Methodology details, benchmark figures, and the specific architectures evaluated are not present in the available source material. Any claim about which axis the survey identifies as most under-served, or which memory mechanism it ranks highest, requires the full document. As a survey rather than an empirical contribution, the paper's analytical weight rests on coverage breadth and the internal consistency of its evaluation criteria—both of which remain to be verified against the underlying text.