Memory System Taxonomy

• Memory System Taxonomy is a framework that categorizes memory mechanisms in both biological and AI systems using cognitive, structural, and operational dimensions.
• It organizes memory into multi-dimensional axes such as object, form, and time, enabling systematic assessment of trade-offs and design choices.
• The taxonomy supports analysis of memory representation, retrieval, and updating processes, influencing architectures from LLMs to hardware systems.

A memory system taxonomy provides a rigorous framework for classifying the mechanisms, architectures, and lifecycle properties through which information is encoded, stored, manipulated, and retrieved in both biological and artificial systems. In the context of modern LLMs, agentic systems, and hardware architectures, memory taxonomies elucidate design trade-offs, operational semantics, and alignment with specific use cases, enabling systematic analysis and principled system design. Contemporary taxonomies are multi-dimensional, encompassing cognitive-inspired categories, engineering axes, structural organization, functional operations, and system-level considerations. This article surveys the principal taxonomic models currently structuring the research landscape.

1. Cognitively-Inspired Axes: Human-to-AI Memory Mapping

Memory systems in LLM-driven AI are frequently mapped onto established human memory categories along axes of time, awareness, and content. The classical distinctions are:

• Sensory Memory: Ultra-brief, modality-specific buffer (iconic, echoic, haptic). AI Analog: Input encoding layers converting raw data into token embeddings/vectors; representations discarded if not processed further.
• Working Memory (Short-Term): Active, manipulable storage (seconds–minutes) supporting reasoning. AI Analog: Non-parametric context windows (prompt buffer, chain-of-thought); parametric KV-caches during inference.
• Episodic Memory (Long-Term Explicit): Temporally contextualized record of personal experiences. AI Analog: Non-parametric long-term memory (external databases or vector stores capturing interaction logs).
• Semantic Memory (Long-Term Explicit): Decontextualized factual and world knowledge. AI Analog: Knowledge embedded in model weights; parameter-efficient finetuning; role-specific adapters.
• Procedural Memory (Long-Term Implicit): Skills/habits, learned through repetition, often inaccessible to conscious recall. AI Analog: Parametric routines, heuristics, and policies encoded directly in model weights, enabling implicit, efficient task execution.

This mapping forms the basis of organizing AI memory features and motivates categorical alignment in system design (Wu et al., 22 Apr 2025).

2. Three-Dimensional and 8-Quadrant Taxonomies

The principal operational taxonomy for LLM-driven AI memory organizes the space along three orthogonal axes (Wu et al., 22 Apr 2025):

• Object ($O$):
  • Personal ($o = P$): User-derived memory (preferences, dialog history).
  • System ($o = S$): Internally generated task/plan artifacts.
• Form ($F$):
  • Non-Parametric ($f = N$): External (vector DB, KV-store, graph).
  • Parametric ($f = P$): Model weights, KV-cache, adapters.
• Time ($T$):
  • Short-Term ($t = S$): Volatile, session-bound context.
  • Long-Term ($t = L$): Persisted, cross-session storage.

The Cartesian product ($\{P,S\} \times \{N,P\} \times \{S,L\}$) yields eight functional "quadrants" denoted $o = P$0, each associated with canonical memory roles and implementations (see table below).

Quadrant	Type	Representative Role	Examples
Q₁ (P,N,S)	Personal/Non-Parametric/Short-Term	Session buffer (Working)	In-session prompt/buffer (ChatGPT, Claude)
Q₂ (P,N,L)	Personal/Non-Parametric/Long-Term	Episodic memory (Long-term)	ChatGPT Memory, MemoryScope, mem0
Q₃ (P,P,S)	Personal/Parametric/Short-Term	Personal cache	Prompt cache, Contextual retrieval
Q₄ (P,P,L)	Personal/Parametric/Long-Term	Personalized fact in weights	Character-LLM, AI-Native Memory
Q₅ (S,N,S)	System/Non-Parametric/Short-Term	System scratchpad	ReAct, RAP, Reflexion
Q₆ (S,N,L)	System/Non-Parametric/Long-Term	Procedural/Workflow history	Voyager, Retroformer, ExpeL
Q₇ (S,P,S)	System/Parametric/Short-Term	System KV-cache	vLLM, FastServe, Orca
Q₈ (S,P,L)	System/Parametric/Long-Term	General knowledge in weights	Memorizing Transformer, Titans, LongMem

This structure enables systematic placement of mechanisms and assessment of coverage/gaps in agentic system functionality (Wu et al., 22 Apr 2025).

3. Structural Taxonomies: Graph, Hierarchy, and Organization

Memory system architectures are further categorized by their structural form, crucially impacting retrieval, update, and reasoning capacity:

• Non-Structural Memory: Flat token buffers, embedding stores, simple key-value logs; optimized for single-hop recall but lack compositional/relational expressivity.
• Structural Memory:
  • Graphs: Memory graphs $o = P$1 organize events/entities as nodes, with labeled edges for semantic, temporal, or causal relations. Enables multi-hop, temporal, and compositional reasoning (Yang et al., 5 Feb 2026).
  • Hierarchies and Trees: Multi-level memory (STM/LTM) with paging and schemas.
  • Hybrid/Composite Objects: MemCube, multi-payload note objects, multi-store backends.

The structural layer is mapped onto the memory lifecycle phases: Extraction, Storage, Retrieval, and Evolution, each admitting task- and architecture-specific variants (e.g., token buffering, triple extraction, graph traversal, schema updates) (Yang et al., 5 Feb 2026, Zhou et al., 23 Jun 2026).

4. Functional and Operational Taxonomies

Agentic and AI memory systems are often decomposed by information representation and lifecycle operations (Du et al., 1 May 2025, Hu et al., 15 Dec 2025):

• Representation Axis:
  • Parametric: Knowledge in model weights ($o = P$2), updated via finetuning or editing.
  • Contextual: Explicit, mutable records (token/vector/graph).
  • Latent: Hidden states, memory tokens, cache embeddings.
• Functional Axis:
  • Factual: User/entity knowledge (profiles, facts, tool schemas).
  • Experiential: Trajectory logs, skill templates, strategic insights.
  • Working: Capacity-limited scratchpads for active computation.
• Key Memory Operations:
  • Consolidation: Migrate transient info to durable storage (weights or external).
  • Indexing: Auxiliary structures for fast retrieval.
  • Updating: In-place modification of records or parameters.
  • Forgetting: Pruning/erasing obsolete or harmful memory.
  • Retrieval: Query matching in context or external stores.
  • Compression: Summarization or dimensionality reduction for resource management.

Mapping research onto this operational lens illuminates, for example, which components are stressed by long-term updating (MemoryBank), by multi-source integration (StructRAG), or by parametric fact-editing (ROME, MEMIT) (Du et al., 1 May 2025).

5. System-Level and Agentic Memory Taxonomies

Recent works emphasize system-oriented classification, reflecting operational bottlenecks and lifecycle requirements (Zhou et al., 23 Jun 2026, Omri et al., 4 Jun 2026):

• Core System Modules (R, S, Q, U):
  1. Representation & Storage: Data model (token, graph, composite) and physical backend (window, vector DB, graph DB, hybrid).
  2. Extraction: Transformation of observations into memory records (raw, semantic, schema-constrained, LLM or tool-mediated).
  3. Retrieval & Routing: Query-to-memory matching using self-attention, dense/sparse retrieval, graph traversals, LLM-generated plans.
  4. Maintenance: Versioning, eviction, consolidation, and conflict/consistency management.

• Construction Pipeline: How history is ingested—absent (pure context), deterministic (BM25/embedRAG), LLM-mediated (GraphRAG, Mem0), or agentic (A-Mem, MIRIX).

• Storage Substrate: Context buffer, inverted index, vector store, graph/multi-view store, hybrid/multi-store.
• Retrieval Pipeline: Passthrough, (dense/sparse) k-NN, graph expansion/rerank, hybrid/agentic-tool-driven, type-routing.
• Mutability: Append-only, consolidation, or free mutation under agentic tool control (Omri et al., 4 Jun 2026).

Empirical system benchmarking and ablation findings indicate that no single configuration dominates across all tasks; trade-offs are context- and workload-dependent (e.g., raw sequences maximize content fidelity but at resource cost, while aggressive compression impairs long-horizon recall). Evaluation must span representation fidelity, retrieval precision/recall, update correctness, and cost/performance curves (Zhou et al., 23 Jun 2026).

6. Memory in Agentic RAG and POMDP Architectures

Autonomous retrieval-augmented architectures formalize memory state ($o = P$3) within a finite-horizon POMDP:

• Working memory ($o = P$4): Short-term scratchpad, appended with observations, retrieved documents, reasoning chains, and pruned as needed.
• Episodic memory ($o = P$5): Session-bounded event logs, vector-indexed.
• Long-term memory ($o = P$6): Persistent vector or knowledge graph across sessions.
• Semantic memory: Static model weights.
• External knowledge store: Not a true agent memory, but the corpus/context queried via explicit agent action (Mishra et al., 7 Mar 2026).

Formally, update and retrieval are defined as: $o = P$7 The agent's policy must coordinate memory module access and evolution, trading off cost, latency, and semantic relevance.

7. Hardware and Descriptor-Based Memory System Taxonomy

At the hardware and operating systems layer, memory system taxonomy addresses scaling, cost, security, and program semantics propagation (Li et al., 5 Aug 2025, Liu et al., 2020, Tong, 31 Oct 2025):

• Volatile vs. Non-Volatile Main Memory (NVMM):
  • Volatile: DRAM
  • NVMM: PCM, STT-RAM, ReRAM, 3D X-Point; classified by architecture (horizontal/hierarchical), persistence management, energy, endurance, programming models.
• Specialized Memory Classes:
  • SRAM, DRAM, StRAM (short-lived, high-endurance, ~1 s retention), LtRAM (read-heavy, persistent, multi-minute scale). Each class matches application-specific access patterns and lifecycles.
• Descriptor-Based Object-Aware Systems:
  • Addressing modes: tagged pointers, shadow tables, fat pointers, hybrid/centroid encodings, in-object descriptors. Systematically analyzed by lookup latency, metadata capacity, ABI compatibility, fragmentation, and hardware complexity (Tong, 31 Oct 2025).

This taxonomy informs co-design of object semantics-aware hierarchies, hybrid persistent/volatile arrangements, and efficient runtime enforcement of spatial/temporal safety.

---

By integrating cognitive, structural, operational, and system-level taxonomies, the modern research landscape is equipped to analyze and systematically extend memory systems for both foundation models and computational hardware. Ongoing research focuses on multimodal integration, policy-driven automation, robust versioning and privacy, agentic control, cost-aware optimization, and alignments with cognitive/biological principles, supporting the development of adaptive and efficient memory-enabled intelligence (Wu et al., 22 Apr 2025, Zhou et al., 23 Jun 2026, Omri et al., 4 Jun 2026, Hu et al., 15 Dec 2025).