Agent Memory Research in 2026: A Data-Driven Survey and Extended Taxonomy
Updated July 28, 2026
ResearchAI & Machine Learningagent-memoryllm-agentssurveytaxonomytemporal-dynamicsmultimodal-memoryneurocognitionopen-accesszenododata-driven
Published on Zenodo: 10.5281/zenodo.20780690 · Companion Dataset: 10.5281/zenodo.20780696 · Repository: github.com/tobias-weiss-ai-xr/agent-memory-research · Interactive UI: browse papers online
What This Is
This extended survey builds on Liu et al.'s foundational "Memory in the Age of AI Agents" (arXiv 2512.13564, ~200 papers up to January 2026) by extending both the paper catalogue and the taxonomy itself. As of July 2026, the curated list covers 1,049 papers — including 595 works published between February and July 2026 (a ninefold increase over the original survey's post-cutoff additions) — revealing cross-cutting themes the original 3×3×3 taxonomy was not designed to capture.
Three Contributions
1. Data-Driven Living Survey Methodology
The paper list is treated as a structured dataset: a single papers.yaml file is the source of truth with 1,049 entries. Automated scripts validate entries, fetch metadata from arXiv and Semantic Scholar, discover new papers via the arXiv API, and generate outputs including README, JSON export, and BibTeX references. A GitHub Actions workflow enforces validation on every contribution.
This methodology is generalisable to any survey that must remain current — it turns updates into a pipeline run rather than manual LaTeX edits.
2. Extended Taxonomy (3 New Dimensions)
The original taxonomy organises papers along three axes — Forms (Token-level / Parametric / Latent), Functions (Factual / Experiential / Working), and Dynamics (Formation / Evolution / Retrieval) — yielding 27 cells. The distribution across the original 3×3 grid is:
| Token-level | Parametric | Latent | |
|---|---|---|---|
| Factual | 456 | 20 | 30 |
| Experiential | 169 | 11 | 15 |
| Working | 283 | 10 | 55 |
We add three orthogonal dimensions:
| Dimension | Levels | What It Captures |
|---|---|---|
| Temporal Dynamics | None → Decay → Consolidation → Bi-temporal | How memory evolves over time: forgetting curves, sleep consolidation, temporal knowledge graphs |
| Modality | Text-only → Multimodal-in → Multimodal-out → Full-multimodal | The sensory channels memory systems handle: vision, language, audio, embodiment |
| Biological Inspiration | None → Cognitive-metaphor → Neuro-inspired → Brain-architecture | Degree of fidelity to human memory: from loose metaphors to detailed neural models |
These dimensions are additive — they enrich rather than replace the original taxonomy, revealing patterns invisible in the 3×3×3 space alone.
The extended taxonomy reveals nine cross-cutting themes: Temporal Dynamics (>130 papers), Multimodality (~100 papers), Graph and Structured Representations (~100 papers), Biologically Inspired (~40 papers), Lifelong Learning (~190 papers, the largest cluster), Benchmarking (>80 benchmarks), Retrieval Fusion (~30 papers), Security & Adversarial Robustness (>120 papers), and Efficiency & Compression (~140 papers).
3. Gap Analysis and Research Roadmap
The analysis identifies:
- 3 persistently sparse cells: Experiential/Latent (15 papers), Working/Parametric (10 papers), Experiential/Parametric (11 papers)
- No unified benchmark — over 80 distinct evaluation protocols, none spanning more than 3 of the 27 taxonomy cells
- 6 open problems: Memory–reasoning trade-off, compositional memory, long-horizon stability, multi-agent sharing, privacy-compliant forgetting, cross-modal transfer
The roadmap spans short-term (unified benchmarks), medium-term (compositional memory, formal trade-off models), and transformative long-term goals (multi-year stability, privacy-compliant forgetting, cross-modal transfer).
Eight Representative Systems
The survey highlights eight systems that illustrate the architectural diversity of the 2026 landscape:
| System | Key Innovation | Cell |
|---|---|---|
| Mem0 | Multi-signal retrieval; 92.5% on LoCoMo | Factual/Token-level |
| FluxMem | Adaptive memory structure selection per query | Factual/Token-level |
| SCM | NREM/REM sleep-inspired consolidation cycles | Experiential/Latent |
| Engram | Bi-temporal knowledge graph with dual-process retrieval | Factual/Latent |
| CraniMem | Neurocognitive gating with bounded episodic buffer | Experiential/Token-level |
| MemRL | Runtime RL on episodic memory | Experiential/Parametric |
| All-Mem | Lifelong topology evolution | Experiential/Latent |
| TeleMem | Long-term multimodal memory for agentic AI | Working/Latent |
Timeline
The field has accelerated dramatically: from roughly 200 papers by January 2026 to 1,049 by July 2026, with 595 new papers in just six months.
Why No arXiv?
The paper was submitted to arXiv in June 2026 but was blocked by the October 2025 review-article policy, which no longer accepts literature reviews. We chose Zenodo as the primary venue (which provides stable DOIs through the CERN infrastructure) and released all data, code, and analysis under an MIT license.
Repository & Interactive UI
The agent-memory-research repository hosts:
papers.yaml— structured source of truth (1,049 entries)scripts/— validation, metadata enrichment, figure generation, BibTeX exportdocs/— interactive GitHub Pages site with filtering by category, date, and keyword searchpaper/— full LaTeX source and compiled PDF
How to Cite
Weiß, T. (2026). Agent Memory Research in 2026: A Data-Driven Survey and Extended Taxonomy (v1.1.0-revised). Zenodo. https://doi.org/10.5281/zenodo.20780690
References
- Liu et al. (2026). Memory in the Age of AI Agents: A Survey. arXiv:2512.13564. https://arxiv.org/abs/2512.13564
- Agent Memory Paper List (original fork): https://github.com/Shichun-Liu/Agent-Memory-Paper-List