Three-Dimensional Taxonomy
The survey organizes agent memory along three orthogonal axes: form (how memory is stored), function (what memory represents), and dynamics (how memory evolves). This visualization shows the conceptual space and highlights real systems positioned within it.
Memory Space: Forms × Functions × Dynamics
Conceptual Boundaries: Agent Memory vs Related Concepts
Memory Forms: Storage Representations
Token-level (explicit text), parametric (neural weights), and latent (continuous vectors) memory representations. Each form has distinct cost, interpretability, and capacity tradeoffs.
Form Comparison Heatmap
Normalized scores across key evaluation dimensions. Darker = better performance in that category.
Token-Level Memory Pipeline
Explicit text storage with retrieval, summarization, and compression stages.
Parametric Memory: Catastrophic Forgetting Over Time
Memory Functions: What Memory Represents
Factual (static knowledge), experiential (past events), and working (temporary state) memory serve different roles in agent reasoning. Real systems blend all three.
Function Distribution in Real Systems
Memory Retrieval Scoring: Generative Agents
Composite score from recency, importance, and relevance. Hover cells to see scores.
Working Memory Hierarchy: MemGPT
OS-inspired memory paging with main context, external storage, and archival tiers.
Cost-Performance Tradeoffs
Production memory systems must balance inference latency, storage cost, update overhead, and interpretability. No single form dominates all dimensions.
Cost Scaling by Memory Form
Pareto Frontier: Latency vs Interpretability
No memory form dominates. Token-level is slow but interpretable; latent is fast but opaque.
Context Window Utilization Over Conversation Length
Real-World Memory Systems
Case studies from research prototypes: Reflexion, Generative Agents, MemGPT, and RET-LLM. Each combines multiple forms and functions.
System Architecture Comparison
Form × Function Matrix Across Systems
Which systems use which combinations? Most are hybrids.
Key References
Memory in the Age of AI Agents (2025)
Shinn et al., 2023
Park et al., 2023
Packer et al., 2023
Modarressi et al., 2024
Dai et al., 2019
Kirkpatrick et al., 2017
Hu et al., 2021
Olsson et al., 2022
Wang et al., 2023
Wu et al., 2022