Memex

• The memex is Vannevar Bush’s 1945 vision of a personal knowledge machine: a device that stores all of one’s books, records and communications and lets them be “consulted with exceeding speed and flexibility” — an intimate, mechanized supplement to memory (Bush 1945). It gives its name to this repository (see README).

• Bush’s central idea is associative indexing: instead of filing an item under one rigid category, you build trails that tie documents together the way a mind jumps by association. The trails between documents are as valuable as the documents themselves. This is hypertext two decades early, and it is exactly the premise of the Zettelkasten: a note derives its quality from the network of links, not from its contents alone.

What Bush couldn’t solve

• The memex was a machine for storage and traversal, but the trails still had to be laid by hand. That is the flaw that kills such systems in practice: the maintenance burden grows faster than the value. People collect but never connect — the collectors fallacy — and the archive rots into an opaque pile.
• Bush had no one to do the bookkeeping. The missing piece was never storage; it was a tireless librarian to keep the trails current.

The realization here

• A large language model is that librarian. It reads sources, writes the summaries, lays and repairs the cross-reference trails, flags contradictions, and keeps the synthesis current — the grunt work that made humans abandon their wikis. Because the cost of maintenance drops to near zero, the vault gets better the more you add to it instead of decaying.
• This flips the usual LLM pattern. RAG re-discovers knowledge from raw documents on every query; the memex compiles it once into an interlinked vault and keeps it maintained, so synthesis and connections compound rather than being re-derived. The human curates sources, directs the analysis, and asks the questions — abstracting up to the level of mental models — while the machine handles everything else. It is a genuine second brain, and a substrate for the creative process of connecting distant ideas, much as AI is reshaping how science itself is done.
• The operating rules for that librarian — how it ingests, answers, and gardens this vault — live in the schema, CLAUDE.md.

Links:

• Zettelkasten — the note-taking method this system automates.
• Large language models — the maintainer.
• Collectors fallacy — the failure mode the memex falls into without a maintainer.
• Andy Matuschak’s working notes (Matuschak n.d.a) — a public vault of densely linked evergreen notes, kept by hand. Its note Prefer associative ontologies to hierarchical taxonomies (Matuschak n.d.b) argues that links, not folders, should carry a note’s place: one note can belong to several fields at once.
• Consensus map of science (Klavans and Boyack 2009) 📄 — twenty maps of science agree on a ring of fields with no centre, not on a tree. The same reason argues for a vault map where no field outranks another.

References:

• Vannevar Bush, As We May Think, The Atlantic Monthly, 1945 (Bush 1945).