Graph Analysis of 282 Agent Memory Research Papers: Network Metrics, Category Shifts, and Publication Velocity
Graph Analysis of 282 Agent Memory Research Papers
Agent memory is one of the fastest-growing areas in LLM research. Between early 2023 and July 2026, the field has expanded from a handful of papers to 282 publications spanning three memory categories and three representation substrates. This article presents a comprehensive graph data science analysis — from network metrics and degree centrality to publication velocity and category mix evolution.
The Dataset
The agent-memory-research repository tracks papers across three functional categories and three representational subcategories:
| Category | Description | Papers | |---|---|---|---| | Factual | Knowledge storage and retrieval | 144 | | Experiential | Insights and skills from experience | 78 | | Working | Active context management | 60 |
| Subcategory | Description | Papers |
|---|---|---|
| Token-level | Explicit, discrete representations | 214 |
| Parametric | Implicit weight-based memory | 26 |
| Latent | Hidden state representations | 42 |
All 282 papers are stored as structured YAML with metadata including title, date, category, subcategory, venue, and URLs — enabling systematic graph analysis.
Methodology
The analysis pipeline works entirely with Python standard library (no PyYAML, no NetworkX). Parse → aggregate → compute → render:
- YAML parsing: State-machine parser extracts 282 records from
papers.yaml - Statistical aggregation: Category/subcategory counts, temporal bucketing, venue frequency, title keyword frequency
- Graph construction: 282 nodes (papers) connected by 496 edges (shared venue + same subcategory-year + same category-year)
- Network analysis: Degree centrality computation, graph density, average degree, hub identification
- Trend analysis: Linear regression on monthly publication counts, year-over-year growth decomposition
- Web rendering: Chart.js for statistical charts, D3.js v7 for the force-directed graph
The entire pipeline runs in less than 1 second and regenerates docs/graph_analysis.html from scratch.
Graph Network Metrics
The force-directed graph yields several network-science metrics that quantify the structure of the research community:
| Metric | Value | Interpretation |
|---|---|---|
| Nodes | 282 | Papers in the corpus |
| Edges | 496 | Relationships (shared venue, subcategory-year, category-year) |
| Graph Density | 0.0125 | ~1.25% of all possible edges exist — a moderately sparse scientific collaboration graph |
| Average Degree | 3.52 | Each paper connects to ~3.5 others on average |
| Maximum Degree | ~28 | Hub papers that bridge multiple venue/category communities |
The low graph density is expected for a bibliometric network spanning 3.5 years. The maximum degree reveals that a small number of papers act as bridges between different research communities — typically survey papers published at major conferences that connect factual, experiential, and working memory subfields.
Force-Directed Graph Analysis
The force-directed graph is the central visualization, where each paper is a node colored by category and sized by degree centrality (how well-connected the paper is):
Node color: Blue = Factual (144)
Green = Experiential (78)
Gold = Working (60)
Node size: Larger = higher degree centrality (more connections)
Smaller = lower degree centrality (fewer connections)
Edge types:
Strong (0.6) = same venue (ACL, ICML, ICLR, NeurIPS)
Medium (0.3) = same subcategory + same year
Weak (0.15) = same category + same year
Key Observations
Degree-centrality sizing reveals hub papers. Unlike the previous recency-based sizing (which penalized earlier work), degree centrality highlights papers that bridge multiple communities — surveys connecting factual and experiential memory, or venue-anchoring papers that co-author-network links radiate from.
Venue clusters dominate the graph structure. Papers published at the same venue form tight clusters — ACL 2026 alone accounts for dozens of connected papers. The force simulation pulls these into distinct spatial groups, revealing which venues are the primary publication channels for agent memory research.
The three memory categories form loose but distinguishable regions. Factual memory papers (blue) dominate the graph center as the largest category. Working memory (gold) and experiential memory (green) tend to cluster in the periphery, often connected through shared-venue edges to the factual core.
Degree distribution is skewed. A small number of survey and taxonomical papers accumulate most of the edges, while the majority of papers have degree 1-3. This power-law-like distribution is typical of citation and collaboration networks.
Category × Subcategory Heatmap
The heatmap reveals where research effort concentrates:
| | Token-level | Parametric | Latent | |---|---|---|---|---| | Factual | 111 | 17 | 16 | | Experiential | 66 | 6 | 6 | | Working | 37 | 3 | 20 |
Token-level representations dominate across all categories (214/282 = 76%). This makes intuitive sense: most agent memory systems store explicit text — conversation history, extracted facts, compressed summaries.
The latent subcategory remains distinctive for working memory (20 vs 16 and 6). This reflects the use of hidden states and recurrent mechanisms for active context management — approaches like MemAgent, Memory-R1, and RMM that learn memory update policies over latent state spaces.
Temporal Trends
Monthly Publication Timeline
The monthly timeline with 3-month rolling average overlay shows three distinct phases:
- 2023 Q2–Q4: Pioneering phase — fewer than 3 papers/month. Foundational work: Generative Agents (Park et al., April 2023), MemGPT (October 2023), MemoryBank (May 2023).
- 2024: Growth phase — 3–8 papers/month. Establishment of benchmarks (LongMemEval, LoCoMo) and taxonomy frameworks.
- 2025–2026 H1: Explosion — 10–25+ papers/month. ACL 2026 alone accepted dozens of agent memory papers across main conference and findings.
Publication Velocity
A dedicated velocity chart plots raw monthly paper counts alongside a 3-month rolling average and a linear regression trend line:
Papers/month ≈ 0.30 × month_index − 0.4
The positive slope confirms sustained acceleration. The 3-month rolling average smooths out conference-driven spikes (e.g., ACL submission deadlines) and reveals the underlying growth rate.
Year-over-Year Growth
| Year | Papers |
|---|---|
| 2023 | ~15 |
| 2024 | ~45 |
| 2025 | ~80 |
| 2026 (through July) | ~140+ |
The field is doubling approximately every 12 months. At current trajectory, the corpus would exceed 400 papers by end of 2027.
Cumulative Growth Curves
The cumulative chart shows all three categories growing in tandem, with factual memory maintaining a consistent lead, followed by experiential memory pulling ahead of working memory in 2026. The curves are approximately exponential (linear on a log scale), indicating sustained growth rather than saturation.
Venue Distribution
Non-empty venues reveal the primary publication channels:
| Venue | Papers |
|---|---|
| ACL 2026 | ~15+ |
| ICML 2026 | ~5 |
| ICLR 2026 | ~4 |
| Findings ACL 2026 | ~4 |
| NeurIPS 2025 | ~3 |
The majority of papers (≈180) have no venue tag (preprints on arXiv), reflecting the field's rapid pace — many papers appear on arXiv months before conference publication.
Subcategory Trends Over Time
A dedicated subcategory trends chart tracks token-level, parametric, and latent representations across the full timeline:
- Token-level has been the dominant approach since 2023, accelerating sharply from 2025 onward — reflecting the explosion of RAG-based and retrieval-augmented memory systems
- Parametric shows sporadic activity, with small clusters around ICML and NeurIPS submission cycles — fine-grained weight-based memory remains a niche approach
- Latent maintains a steady but lower baseline, with notable spikes around working memory papers that use hidden-state policies
The three subcategories grow in concert but at very different scales, reflecting the heavy token-level bias in current research infrastructure (vector databases, embedding models, RAG pipelines).
Category Mix Evolution
A new per-year donut chart series reveals how the research focus has shifted over time:
- 2023: Balanced mix — factual (40%), experiential (30%), working (30%). The pioneering phase had no dominant paradigm.
- 2024: Factual pulls ahead (~50%) as benchmark-driven knowledge retrieval research ramps up.
- 2025: Experiential begins to close the gap (~30%) as reflection-based and skill-learning memory systems gain traction.
- 2026 (through July): Experiential surges to ~35%, narrowing the gap with factual (~50%). Working memory stays at ~20%, though the absolute count is growing.
This shift from "what agents know" (factual) to "what agents learn from experience" (experiential) reflects the maturation of the field — from static knowledge bases toward dynamic, self-improving memory systems.
Title Keyword Frequency
A new keyword-frequency analysis extracts the 30 most common words from all 282 paper titles:
- Top words: Memory, Agent, LLM, Learning, Retrieval, Long-term, Context, Knowledge, Reasoning, Framework
- Emerging terms in 2026: Self-evolving, Adaptive, Personalized, Tool, Multimodal, Continual
- Declining terms: RAG (becoming infrastructure, not a differentiator), Chat (giving way to Agent)
The keyword distribution confirms the field's trajectory: from storage-oriented "memory for chatbots" toward action-oriented "adaptive memory for autonomous agents."
Radar Chart: Category Composition by Subcategory
The radar chart compares the three subcategories across the three functional categories:
- Token-level dominates on all axes (96 factual, 54 experiential, 43 working)
- Parametric is the smallest across all categories (16, 6, 3)
- Latent shows a distinctive spike for working memory (27 vs 11 and 3)
This asymmetry is the most interesting structural pattern: while factual and experiential memory rely overwhelmingly on token-level representations, working memory has a significant latent-state subcommunity.
Interactive Analysis Tool
All charts are available in the interactive analysis page:
- Force-directed graph: Drag nodes, zoom, hover for paper details, click to open paper URL
- Timeline with moving averages: Toggle series visibility by clicking legend items
- Stacked timeline: Category composition per month
- Year-over-year bar chart: Clear annual growth trajectory
- Venue distribution: Horizontal bar chart of publication venues
- Radar chart: Subcategory composition per category
- Category × Subcategory heatmap: Canvas-rendered matrix
Code and Reproducibility
The analysis pipeline is open source:
- Repository: github.com/tobias-weiss-ai-xr/agent-memory-research
- Analysis script:
scripts/generate_analysis.py(pure Python stdlib, under 1 second runtime) - Data format:
papers.yamlwith 259 entries and 10 metadata fields each - Output:
docs/graph_analysis.html— self-contained HTML with embedded Chart.js and D3.js
Key Takeaways
- Agent memory is a high-growth field — doubling annually, now at 259 papers
- Token-level representations dominate (74.5% of papers), but latent-state working memory is a distinct subcommunity
- ACL 2026 is the dominant publication venue, reflecting the NLP community's embrace of agent memory as a core problem
- Venue-based graph clustering reveals research communities — the force-directed graph shows clear separation between conference communities
- Factual memory leads by volume, but working memory is catching up fast as RL-based memory policies gain traction