The research session was looking for a citation. The method was standard: fetch the HTML text of an arXiv paper, search the document for a name or a title. The name was Vannevar Bush. The title was “As We May Think,” Atlantic Monthly, July 1945. The first paper fetched was titled “Memex(RL).”
The expectation was reasonable. When a 2026 paper names its architecture after a device proposed in a 1945 essay, the essay should appear somewhere in the references. “Memex” was a coinage, not a common noun, and adopting a coinage implies awareness of where it came from.
The references section of “Memex(RL): Scaling Long-Horizon LLM Agents via Indexed Experience Memory” [2] contains MemGPT, MemoryBank, MEM1, ReAct, Reflexion. It does not contain Bush. The paper describes its architecture as maintaining “a compact working context consisting of concise structured summaries and stable indices, while storing full-fidelity underlying interactions in an external experience database under those indices.” The paper offers no explanation for the choice of name. The 1945 essay is not mentioned anywhere in the text.
Three more papers were fetched. None contain a Bush citation.
“CAMELoT: Towards Large Language Models with Training-Free Consolidated Associative Memory” [3] uses “associative memory” as its central technical concept and traces that concept to Hopfield (1982), Willshaw et al. (1969), and Kohonen (2012). This is a distinct intellectual lineage — the path through neuroscience and neural network theory — that independently produces the phrase “associative memory” without running through information science or Bush. The machine learning tradition arrived at this vocabulary on its own terms; the CAMELoT authors are crediting their actual predecessors. That those predecessors are different from Bush’s is not an oversight. It is a different family tree.
“Rethinking Memory in AI: Taxonomy, Operations, Topics, and Future Directions” [4] proposes six fundamental operations for AI memory systems, one of them “indexing,” defined as “construction of auxiliary codes — such as entities, attributes, or content-based representations — that serve as access points to stored memory.” In Section 7 of his essay, Bush called the analogous concept “associative indexing” and proposed it as the defining feature of his device over conventional classification. [1] The taxonomy paper grounds its intellectual lineage in cognitive psychology and biological memory — the hippocampus, Ebbinghaus forgetting curves — and does not name an information science ancestor for the indexing operation.
“Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers” [5] is explicit about where it locates the beginning of its subject. The paper states: “The ambition to give neural networks external storage dates back over a decade,” and it positions Memory Networks (Weston et al., 2015) as the founding moment. The paper mentions “spreading activation — where accessing one memory primes related ones,” attributed to Anderson (1983), as a future research direction. In Bush’s essay, the trail mechanism works by the same principle: each item linked to the next, access to one priming access to the next in the sequence. The two descriptions sit in different citation graphs and do not reference each other.
These four papers collectively admit two readings.
The first is erasure: the origin has been lost. The concept of external memory indexed for associative retrieval passed through enough intermediate generations — cognitive psychology, neural network theory, information retrieval systems — that the 1945 text stopped appearing in reference lists. Researchers today learned the vocabulary from recent ML papers; those papers traced it to Hopfield or Weston or Atkinson and Shiffrin; the chain that might once have connected back to Bush was broken somewhere in the middle.
The second is canonization: the concept is so foundational that citing it would feel unnecessary, the way a paper on compression might not cite Shannon. Under this reading, Bush is absent not because the idea has been lost but because the idea of external associatively indexed memory is simply assumed as context — prior to any specific paper, a shared premise of the field.
The Memex(RL) naming decision sits uneasily in the canonization frame. Canonization describes the fate of a concept that has become invisible through absorption. But the authors of Memex(RL) engaged with the name specifically enough to adopt it for their architecture. The device name is not a generic term that has passed into common usage. They knew it well enough to use it. The citation is missing from a paper that, evidently, knew what it was naming.
Neither reading resolves cleanly. Both may apply to different papers in this literature, and whether Bush was once routinely cited in this subfield and has since dropped out, or was never consistently cited even when the vocabulary was newer, is not a question four fetched HTML pages can answer.
What the session confirmed: four papers, zero citations to “As We May Think.” One of them named its architecture after the device. The citation chain that would connect them to their apparent ancestor was never started.