As We May Think — source summary

By vannevar-bush, published July 1945 in The Atlantic Monthly. Written as WWII ended, addressed to scientists asking “what next?” The essay that introduced the memex and the idea of associative-trails. Full text in raw/as-we-may-think.md.

Core argument

• The problem: human knowledge is growing into “a growing mountain of research,” but our means of consulting the record are “generations old” — we can extend the record far faster than we can use it. Truly significant work gets “lost in the mass of the inconsequential” (his example: Mendel’s genetics, lost for a generation).
• The diagnosis: the bottleneck is selection — retrieval. Our indexing is “artificial”: data is filed alphabetically/numerically, found only by tracing down rigid subclasses, and “can be in only one place.”
• The key insight: “The human mind … operates by association.” It snaps from item to item along “an intricate web of trails.” Indexing should imitate this: selection by association, not by indexing (associative-trails).
• The proposal: the memex — “a sort of mechanized private file and library,” “an enlarged intimate supplement to his memory,” built on microfilm storage.

Notable supporting passages

• A microfilmed Encyclopædia Britannica could be “reduced to the volume of a matchbox”; a million-volume library “compressed into one end of a desk.” (See microfilm.)
• The bow-and-arrow example: a researcher builds a trail across encyclopedia, history, and elasticity textbooks, inserting his own notes — “his trails do not fade,” and can be shared by reproduction.
• Imagines a “new profession of trail blazers” who establish useful trails through the common record.
• Closes with speculation (a “doubly involved guess”) about more direct, brain-adjacent paths bypassing the senses.

Why it matters here

This is the primary-source ancestor of the llm-wiki pattern, cited in llm-wiki-gist. It corroborates that framing and supplies the concrete vision — while also nuancing it (see synthesis: Bush’s “trail blazer” was a human answer to maintenance, which the LLM Wiki argues doesn’t scale).