- [2026/01/15] 🚀 Project Website: https://baijia.online/homepage/memory_survey.html
- [2026/01/15] 🚀 Full survey released! We establish a unified framework for AI memory mechanisms.
- [2026/01/07] 🚀 Project repository initialized.
This survey presents a comprehensive overview of AI memory mechanisms anchored in a unified theoretical framework. We propose a structured "4W Memory Taxonomy" to enable consistent analysis across diverse architectures. Building on this foundation, we systematically review memory systems in both single- and multi-agent contexts, examining their architectures, functions, applications, and evaluation methodologies. By synthesizing cognitive theories with engineering benchmarks, this work provides a coherent roadmap for advancing the theoretical understanding and technological development of AI memory.
To clarify the scope of AI memory, we distinguish between three interrelated layers:
- LLM Memory: The low-level computational kernel for prediction, consisting of Parametric Weights (static) and Context Window (runtime).
- Agent Memory: The functional workflow supporting autonomous operation and complex task execution via perception-planning-action loops.
- AI Memory: The overarching cognitive concept aimed at lifelong evolution, long-term persistence, and adaptation.
We establish a structured 4W Memory Taxonomy to enable consistent analysis across diverse architectures:
- When (Lifecycle Dimension): Examines the temporal span of memory, including Transient (extremely short-lived), Session (task duration), and Persistent (cross-session retention).
- What (Type Dimension): Categorizes by the nature of stored information, including Procedural (skills), Declarative (facts), Metacognitive (reflections), and Social/Personalized (user models).
- How (Storage Dimension): Explores technical implementation, from Implicit Storage (Parametric/Latent) within model weights to Explicit Storage (Raw Text, Vector DB, or Structured Graphs) outside the model.
- Which (Modality Dimension): Classifies by information formats, covering Single-modal (text-only) and Multimodal (fusing images, audio, and video).
Effective collaboration within MAS hinges on communication mediated by memory sharing. We organize these mechanisms into two core dimensions:
- Communication Mechanisms: Ranges from Explicit Communication (interpretable symbols like natural language or structured schemas) to Implicit Communication (latent representations/hidden embeddings).
- Memory Sharing Mechanisms: Categorized into Task-Level (experience accumulation and knowledge transfer) and Step-Level (precise context allocation and role-aware filtering).
We categorize memory evaluation into four essential dimensions to provide a structured assessment of agent memory:
If you find this survey or the established taxonomy helpful in your research, please cite our work coming soon