Agent Memory Research in 2026: A Data-Driven Survey and Extended Taxonomy

Agent memory has emerged as a first-class primitive in the pursuit of intelligent, long-horizon AI agents. The original survey by Liu et al. established a unified 3x3x3 taxonomy—Forms x Functions x Dynamics—organizing roughly 200 papers up to January 2026. Since then, the field has accelerated dramatically: our living catalog now spans 952 papers (as of July 2026), including 498 works published since February 2026—an eightfold increase over the original survey's post-cutoff additions. This explosive growth reveals cross-cutting themes the original taxonomy was not designed to capture. In this paper we make three contributions. First, we describe a fully data-driven methodology for curating and maintaining a living paper survey, with a single-source-of-truth YAML file, automated validation pipelines, bulk metadata enrichment, and an open-source infrastructure. Second, we extend the original taxonomy with three novel orthogonal dimensions—Temporal Dynamics, Modality, and Biological Inspiration—that capture the emerging themes of forgetting curves, multimodal memory, and neurocognitively inspired architectures, and we document two further theme clusters—security and adversarial robustness, and efficiency-driven compression—that have grown into major research strands in their own right. Third, we conduct a gap analysis that identifies persistently underpopulated taxonomy cells, missing evaluation standards, and a concrete research roadmap spanning short, medium, and long-term milestones. The nine cross-cutting themes identified are: (1) Temporal Dynamics (130+ papers), (2) Multimodality (~100 papers), (3) Graph and Structured Representations (~100 papers), (4) Biologically and Neurocognitively Inspired (~40 papers), (5) Lifelong and Continual Learning (~190 papers, the largest cluster), (6) Benchmark and Evaluation (80+ papers), (7) Retrieval Fusion (~30 papers), (8) Security and Adversarial Robustness (120+ papers), and (9) Efficiency and Compression (~140 papers). All data, code, and a browsable web interface accompany this paper at github.com/tobias-weiss-ai-xr/agent-memory-research.