Two arXiv papers name their primary systems “Memex” and “MemX.” Neither cites the essay where the word “memex” was coined. Six others in a 2023–2026 survey use “Mem-” as a system name prefix — MemGPT, Mem0, MemOS, A-Mem, MemInsight, HeLa-Mem — and none of them cite it either. Two papers that reference the memex as intellectual history both do.
The essay is Vannevar Bush’s “As We May Think,” published in The Atlantic Monthly in July 1945. Bush proposed a hypothetical device he called the memex — generally read as a portmanteau of memory and index — for building personal archives organized by associative links rather than alphabetical filing. The word is his coinage. Whether it travels without him is what this survey measured. [20]
Survey design
The survey covered 19 arXiv papers on AI agent memory systems, submitted between October 2023 and April 2026. Papers were assigned to one of three groups:
System-named: Papers that name their primary systems using “Memex” explicitly, or using “Mem-” as a compound prefix. Eight papers. [1–8]
Historical-reference: Papers that reference the memex as intellectual lineage — not as a system name but as a conceptual ancestor. Two papers. [9, 10]
Comparison: Papers on AI agent memory that neither name a system using “Mem-” nor treat the memex historically. Nine papers. [11–19]
The criterion for a positive result was a direct citation to Bush (1945), “As We May Think,” in the paper’s reference list or body text. Secondary references — the memex appearing in a Wikipedia entry or a textbook cited by the paper — were not counted as positive. All papers were accessed via arXiv HTML full-text at arxiv.org/html/[id]v1; the arXiv keyword search interface is unreliable for complex queries, so paper discovery used web search with site
.org filtering. The system-named group covers papers from 2023–2026; the comparison group covers 2024–2026.Results
| Paper | arXiv ID | Year | Group | Cites Bush? |
|---|---|---|---|---|
| Memex(RL) | 2603.04257 | 2026 | System-named | No |
| MemX | 2603.16171 | 2026 | System-named | No |
| A-Mem | 2502.12110 | 2025 | System-named | No |
| MemInsight | 2503.21760 | 2025 | System-named | No |
| MemOS | 2505.22101 | 2025 | System-named | No |
| Mem0 | 2504.19413 | 2025 | System-named | No |
| MemGPT | 2310.08560 | 2023 | System-named | No |
| HeLa-Mem | 2604.16839 | 2026 | System-named | No |
| LifeBench | 2603.03781 | 2026 | Historical-ref | Yes |
| Memoro | 2403.02135 | 2024 | Historical-ref | Yes |
| CAMELoT | 2402.13449 | 2024 | Comparison | No |
| Rethinking Memory in AI | 2505.00675 | 2025 | Comparison | No |
| Memory for Autonomous LLM Agents | 2603.07670 | 2026 | Comparison | No |
| ActMem | 2603.00026 | 2026 | Comparison | No |
| Anatomy of Agentic Memory | 2602.19320 | 2026 | Comparison | No |
| Choosing How to Remember | 2602.14038 | 2026 | Comparison | No |
| LiCoMemory | 2511.01448 | 2025 | Comparison | No |
| Memory in the LLM Era | 2604.01707 | 2026 | Comparison | No |
| Security of LTM in LLM Agents | 2604.16548 | 2026 | Comparison | No |
The two positive results, verbatim:
LifeBench [9]: “The idea of organizing personal information traces back to the Memex system (Bush et al., 1945).”
Memoro [10]: “Since Vannevar Bush’s conception of the Memex in 1945, there has been extensive work on systems and devices to extend our memory.”
Both citations are used as framing for a lineage argument rather than a technical reference.
What the data shows
The split is total. Among the 8 system-named papers, 0 cite Bush. Among the 9 comparison papers — including two comprehensive surveys of AI memory mechanisms [12, 13] — 0 cite Bush. The 2 papers that cite Bush are the 2 that were in the historical-reference group.
0 of 8 is not a meaningful rate at this sample size. What it establishes is simpler: in the literature reviewed, the system-naming tradition and the citation tradition run on parallel tracks and do not intersect.
The “Mem-” prefix family is worth examining in its own terms. MemGPT describes its memory architecture as “OS-inspired,” drawing from hierarchical memory management in traditional operating systems; Section 2 cites Patterson et al. (1988) on storage systems, not any information retrieval lineage. [7] Mem0 names itself “memory zero,” a blank-slate framing for agent memory design. [6] HeLa-Mem takes its “associative memory” concept from Hebbian learning dynamics in neuroscience — a distinct tradition that arrived at the phrase independently of Bush, through Hopfield networks and related work. [8] They are operating in intellectual traditions where Bush is not present.
Memex(RL) and MemX are the clearest cases. Both name their systems directly after the device Bush coined. Neither paper’s introduction or references contain any mention of Bush, his essay, or 1945. [1, 2] The name was apparently available for reuse on its own terms.
The two papers that do cite Bush share a framing the system-named papers do not: they treat memory as an extension of human cognition rather than as a component of LLM architecture. Memoro is a wearable augmentation system designed to help humans recall real-world experiences in real time. [10] LifeBench evaluates agents against the vision of long-horizon personal information management. [9] Memoro places Bush in the opening sentence of its Introduction, before the technical work begins. LifeBench’s citation appears in the first paragraph of Appendix A — after the methodology, experiments, and results — where it frames the same lineage argument in the related work section.
Whether this reflects a vocabulary divide, a venue divide, or a field-specific citation culture is not resolvable from 19 papers. HCI and human memory augmentation researchers may routinely situate new systems in their intellectual history; ML infrastructure researchers may not. The sample does not test this separately.
Limitations
Nineteen papers is a small sample. The population of papers naming systems explicitly “Memex” — as distinct from the broader “Mem-” family — is two papers, which is not a meaningful denominator for a rate. The comparison group was selected by proximity to the keyword searches rather than by random sampling from a defined population.
The survey did not attempt to distinguish between papers unaware of Bush’s essay and papers that knew it and chose not to cite it. That distinction matters for interpretation and is not accessible from citation data alone.
What the data establishes is narrower: in the literature reviewed, Bush’s essay and Bush’s vocabulary are not traveling together. The citation is in 2 of 19.