As We May Think

TL;DR: In 1945 Vannevar Bush described the memex: a desk that stores everything you've ever read, written, or recorded, and lets you navigate it not by index but by associative trails — persistent, annotated, shareable paths through your own knowledge. Storage, retrieval, and content addressing were all off by many orders of magnitude in 1945. Today all three exist, and most of the memex has quietly been built — except its most original component, the trail as a first-class, provenance-bearing artifact you can hand to someone else. That piece is the interesting one, and it is now cheap to build.

The idea as Bush had it

Bush, fresh from running the wartime OSRD, argued that science had multiplied our muscles and senses but not our memory. The specific failure he identified is still the right one: the bottleneck of knowledge work is not acquisition but re-finding — "the artificiality of systems of indexing." Libraries file alphabetically or hierarchically; you drill down a taxonomy, find one item, and climb back out. The mind, he observed, does not work that way. It works by association: one item snaps to the next along paths carved by use.

So he proposed a machine that operates the way the mind does. The memex is a desk with translucent screens, a keyboard, and levers. Inside: your books, papers, correspondence, photographs, and notes, stored on microfilm at extreme reduction — Bush estimated the Encyclopaedia Britannica could shrink to the volume of a matchbox, a million-book library into part of the desk. You buy content on film, and you add your own via "dry photography" (a document camera on the desktop) and voice notes.

The navigation primitive is the trail. Viewing two items side by side, you tap a key and permanently join them under a code word. Later, invoking the trail replays the sequence: item, annotation, item, side-branch, item. Trails don't fade like human associations; they can be copied, inserted into other trails, and — crucially — given to other people. Bush imagines physicians consulting a colleague's trail through the literature on a diagnosis, lawyers inheriting annotated trails through case law, and a new profession of "trail blazers" whose product is not documents but paths through documents.

Two things make this more than a fancy filing cabinet. First, links are made by the reader, not the author — the opposite of today's web, where links are authorial and one-directional. Second, the trail is an object: it persists, carries the annotations and the ordering, and is detachable from its creator.

Why it could not work in 1945

Every layer of the stack was missing, mostly by factors of a million or more.

Storage. Microfilm could plausibly hold the data — a million books at roughly a megabyte of text each is on the order of a terabyte-equivalent, and Bush's density estimates weren't crazy. But microfilm is write-once, sequential, and analog. There is no update-in-place, no copy-in-milliseconds, and "inserting an item into a trail" means physically duplicating frames.

Retrieval. Bush had actually built the closest thing that existed: his 1930s "rapid selector" at MIT spun microfilm reels past photocells that matched pre-punched index codes — mechanical grep over holes, at reel-scan speeds. Selection was by exact code only. Access to an arbitrary record meant seconds to minutes of mechanical winding; a modern ANN query over a lifetime's embeddings runs in single-digit milliseconds. That's roughly a 10^4–10^6\times gap in latency, and an unbounded gap in what can be matched.

Content addressing. This is the fatal one. Trails as Bush described them require the machine to know only codes the user assigns. But the vision — machine-assisted association, speech notes, the speculated "machines which will select by association rather than indexing" — needs the machine to understand content. In 1945 there was no OCR, no speech recognition beyond Bell Labs' Vocoder-style analysis, no digital computers in service (ENIAC was announced in 1946), and no theory of semantic similarity. Bush could specify the interface but not the mechanism; the mechanism required representations of meaning, which took another 70 years.

What changed

Three curves crossed the threshold.

Storage is free. A lifetime of everything you read, write, hear, and see — text, screenshots, compressed audio — fits comfortably in a few terabytes; add continuous compressed video and it's tens of terabytes. At roughly $10–20 per terabyte of disk, a whole-life archive costs less than the desk Bush would have put it in, and it fits locally, which matters for the personal character of the memex.

Association is a primitive. Embedding models give a geometric meaning to "this reminds me of that": items are points, association is proximity, and approximate nearest-neighbor indexes make the lookup effectively instant at billion-item scale. This is the mechanism Bush could only gesture at — selection by association rather than indexing, literally.

Traversal has an operator. LLMs can read a trail, summarize it, extend it, and — importantly — propose trails: given a question, walk the archive, assemble an ordered, annotated path, and show its work. Bush's "trail blazer" profession can be partially automated, with the human as editor rather than laborer.

What a serious 2026 revival looks like

Most "second brain" products rebuild the storage and search layers. The unbuilt part is the trail as an artifact. A faithful revival would look like this:

Substrate. Local-first capture of documents, mail, browsing, screenshots, and (opt-in) ambient audio, stored append-only with content hashes. Every item gets multimodal embeddings; the index lives on-device. Nothing here is research anymore.

Trails as data. Define a trail as a signed, versioned object: an ordered list of (content-hash, annotation, timestamp, join-points to other trails). Because items are referenced by hash, a shared trail carries verifiable provenance — you can check that the quoted passage is really the passage, and see exactly what the trail's author added. This is Ted Nelson's transclusion done with modern content addressing, scoped down to something buildable in a quarter.

LLM as trail apprentice. Two modes. Recording: as you work, the model watches the sequence of items you touch and offers to crystallize it into a trail with generated annotations, which you edit — capture must be near-zero-effort or trails won't exist, which is arguably why Bush-style trails never took off in hypertext systems. Blazing: you ask a question and the model returns not an answer but a trail — a path through your own archive and cited external sources, replayable and shareable. This reframes RAG: retrieval traces are usually transient scaffolding thrown away after generation; here the trace is the product.

Exchange. Trails are files. You send a trail; your colleague's memex resolves the hashes it has, fetches or requests what it lacks, and can splice your trail into theirs — Bush's physician-consults-a-trail scenario, verbatim.

Keep from the paper: the trail data model, reader-made links, human-in-the-loop curation, the insistence on personal ownership of the archive. Replace: microfilm with hashes and embeddings, levers with language, and manual code words with learned association — while keeping the user's explicit joins as ground truth that the learned associations must respect.

Descendants, vindication, and what's still open

The memex is the most-vindicated speculative paper in computing, and the lineage is explicit: Engelbart credited it as the seed of NLS and the 1968 "mother of all demos"; Nelson's Xanadu was an attempt to build the trail/transclusion layer properly; the Web shipped the degenerate version — authorial, one-way links, no trails, no provenance. Gordon Bell's MyLifeBits at Microsoft (2000s) was a deliberate memex reconstruction and found that capture was easy but use was hard — exactly the gap embeddings and LLMs have since closed. Recent screen-recording memory products (Rewind, Microsoft's Recall) rebuild the capture-and-search layer; RAG over personal corpora rebuilds associative retrieval. None of them treats trails as persistent, shareable, provenance-bearing objects. Seventy-nine years on, that idea is still sitting on the table.

Open questions that are genuinely research-shaped: whether LLM-proposed trails match the ones expert humans would blaze (an evaluation problem with no benchmark); how to merge trails from different people whose archives overlap only partially; and whether an economy of trails — Bush's trail-blazer profession — has viable incentives, which is a marketplace-design question, not an ML one.

Where to read it

The essay is short, non-technical, and better written than almost anything it inspired: "As We May Think," The Atlantic Monthly, July 1945 (ACM DL copy; bibliographic details verified). Read it alongside Engelbart's "Augmenting Human Intellect" (1962), which turns Bush's sketch into a research program; Nelson's Literary Machines for the maximalist version of trails and transclusion; and Gemmell, Bell & Lueder's MyLifeBits papers (c. 2002–2006) for an honest report of what happened when someone actually tried it two decades too early.