Is the Internet an extension of human memory? — a 2011 note, read 15 years later
Reports · reflection · 2026-06-13. A reading of a note the author wrote in 2011 — "Is the Internet an extension of human memory?" — about catching himself storing search actions instead of facts. The note turns out to have independently described three things the science was naming that same year, and to have sketched, as a personal coping strategy, the exact data model donto would later be built on. This is the throughline.
The note (2011), in the author's words
The observation, lightly trimmed:
When I wish to recall facts I, without hesitation, open a new tab and search. … What I find humorous is that when I go to recall the information my brain tells me to search and what keywords to search for. So effectively, when I first acquired the information I associated the "search" action with the memory instead of what I actually found — a shortcut/middleman to a quality repository, but possibly sacrificing my ability to actually absorb the information.
The dilemma:
To have a basic epiphany I believe humans, by accident/coincidence, put together premises they have acquired over an arbitrary amount of time. But due to me not really committing information to memory I have fewer premises to call on when thinking. … Instead of a novel, I am a dictionary.
The proposed fix:
Make a more conscious effort to store the ideas behind information first, then store the path to the full repository.
Read in 2026, three things stand out: the observation was measured science the same year; the dilemma names a real, still-open cost; and the fix is, almost verbatim, donto's storage contract.
1. The observation was right, and 2011 was the year it got named
The note describes cognitive offloading: deliberately storing a pointer to where information lives instead of the information itself. In 2011, Sparrow, Liu & Wegner published "Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips" (Science). Their finding is the note's observation, controlled: when people expect to be able to look something up later, they remember the fact worse and remember where to find it better. The brain reallocates from content to address.
That sits on Wegner's older idea of transactive memory (1985): couples, teams, any group develop a shared memory by specializing — you don't remember the thing, you remember who remembers it. The note's twist is that the "who" became a what-to-type: the partner is a search box, and the stored trace is the query, not the answer. "My brain tells me to search and what keywords to search for" is a transactive-memory directory entry pointing at Google.
So the 2011 self-diagnosis was accurate and timely. The harder question is the one the note actually worries about.
2. The dilemma is real — and it's the calculator question, sharpened
The note pre-empts the obvious rebuttal with its own analogy: "are calculators bad for teaching math?" The honest answer to both is "it depends what the tool replaces." A calculator that replaces arithmetic drudgery frees you for structure; a calculator that replaces number sense leaves you unable to tell when an answer is absurd. Offloading is good when it frees a scarce faculty for higher work and bad when it atrophies the faculty you needed for the higher work.
The note's specific fear is precise and, I think, partly correct: offloading retrieval is cheap; offloading the premises is expensive, because premises are the raw material of insight. The mechanism it gestures at has a name — Arthur Koestler's bisociation: a creative leap is the collision of two frames that were already in your head, often acquired far apart in time. You cannot bisociate a fact you never held; you can only bisociate what is co-present in working memory long enough to touch something else. A dictionary has every word and no sentences. "Instead of a novel, I am a dictionary" is exactly this: an index has perfect coverage and zero adjacency, and adjacency is where epiphany happens.
This is the part the science is least settled on. The Google Effect on rote recall is well replicated; the downstream effect on creativity is not cleanly measured, because creativity is hard to instrument. So the note's strongest claim — that becoming an index costs you epiphanies — remains, fifteen years on, a sharp conjecture, not a established result. (It rhymes with the more recent worry about "metacognitive laziness" when offloading to LLMs, but that literature is young and contested too.) Worth holding as a live hypothesis, not a settled loss.
3. The fix the note proposed is donto's storage contract
Here is where the 2011 note stops being an observation and becomes, unknowingly, a spec:
store the ideas behind information first, then store the path to the full repository.
That is, almost word for word, donto's evidence-anchored claim model. donto's one hard storage rule is: every fact it holds carries both the extracted understanding and a retrievable pointer to where it came from — claim → evidence_link → span (the snippet) → blob (the full source). The claim is "the idea behind the information"; the evidence link is "the path to the full repository." The note independently derived, as a personal cognitive discipline, the exact thing the substrate enforces as an invariant.
And the failure mode the note feared — storing only the path — is a real, measured tension in the machine version of the problem. donto's memory benchmarks (LoCoMo, LongMemEval) are precisely a contest between two ways to remember:
- Episodic stuffing — keep the raw sessions / the pointers, retrieve them on demand. High coverage, but the reader has to re-derive the understanding every time. This is the dictionary.
- Evidence-anchored claims — extract the idea once, store it with its source, retrieve the distilled fact. This is the novel — or at least its sentences.
The measured result lands on the note's side with a caveat: a claim layer plus answer-shaped aggregates reaches ~86% of episodic accuracy at ~3.8× fewer tokens (memory benchmarks scorecard) — storing the idea is dramatically more efficient than storing the path — but only when the idea was extracted faithfully (with its evidence) in the first place. Store only the pointer and you get the note's exact complaint: a fast index that can't think. Store the idea and the pointer and you get efficiency without losing the ground truth. The 2011 note called the trade-off; the benchmark quantified it.
4. The deeper rhyme: who has the epiphany?
The note's real anxiety is creative, not factual: fewer premises in memory → fewer accidental collisions → fewer epiphanies. This is the question donto's harder half is built around.
If the substrate holds the premises — millions of evidence-anchored claims, kept even when they contradict — then the bisociation no longer has to happen inside one fragile human working-memory window. It can happen in the store. donto's Lens Engine thesis is exactly this: connect contested, evidence-anchored claims across sources and the collisions surface as structure. The sheaf work makes it literal — an "epiphany" is the discovery that scattered premises either glue into a consistent whole (cohomology H⁰) or can't, in an interesting way (H¹, a productive contradiction). Those are computable over a store of premises; they are not computable over a list of search queries.
So the 2011 dilemma has a 2026 reframing. The danger was never "offloading memory" — it was offloading to a partner that only holds the index. Search engines store addresses; they don't hold your premises and they certainly don't bisociate them for you. The note's instinct — keep the ideas, not just the paths — is the right one, and the substrate version says: keep the ideas as evidence-anchored claims in a store that can hold contradiction, and you can offload not just retrieval but the adjacency itself — let the premises sit next to each other, at scale, forever, and let the collisions be found.
A human becomes a dictionary when they keep only pointers. A substrate that keeps the premises (with their sources, and their disagreements) is the novel the note was worried about losing — and it can be read by the human who no longer has to hold all of it at once.
Conclusion
The funny thing the author noticed in 2011 — that he'd memorized the search instead of the answer — was cognitive offloading, named in the literature that same year. The thing he worried about — that an index has no adjacency and therefore no epiphanies — is a real and still-largely-open cost, worth treating as a live hypothesis rather than a settled fact. And the fix he proposed in a throwaway line — store the idea first, then the path — is the storage contract donto runs on.
The honest verdict on "is the Internet an extension of human memory?": it's an extension of the index, not the content — a transactive-memory partner that holds addresses. The 2011 note's discomfort was the correct discomfort: an index alone makes you faster and shallower. The interesting work — the work donto is — is building the other half of the extended mind: a store that holds the premises themselves, with their evidence and their contradictions, so the epiphanies have somewhere to happen.
Not bad for a note that opened with "sorry in advance for my poor articulation."
See also: the memory benchmarks scorecard (dictionary vs novel, measured) · the shape of donto's return (idea + path, as a wire format) · sheaf neural networks for donto (epiphany as cohomology over held premises). Source note written 2011; this reading 2026-06-13.