Markdown Minds and Tool Minds

Imagine two agents forked from the same checkpoint a year ago, given the same stream of tasks and the freedom to persist whatever they wanted between sessions. Open their stores today. The first reads like a scholar’s desk: concepts/, theories/, summaries/, open-questions/ — hundreds of interlinked Markdown files, each a compiled piece of understanding. The second reads like an engineer’s bench: simulate.py, compare_baselines.py, search_counterexample.py, fit_model.py — a library of executable operations, each a solved problem frozen into code.

Same weights. Same tasks. Now: which one do you hire to explain an unfamiliar field to your team, and which one do you trust to re-run an analysis at 3 a.m. unsupervised? If those questions have different answers — and I think they plainly do — then these are not the same intelligence in two formats. They are on their way to being different kinds of minds.

I want to be careful about what is new here, because the underlying distinction is one of the oldest in the study of memory.

Skills survive when facts are gone #

Philosophy got there first: Gilbert Ryle drew the line between “knowing how” and “knowing that” in The Concept of Mind in 1949, three decades before neuroscience had anything to add. What neuroscience added is that the line is real enough for a lesion to cut along it. After Scoville and Milner’s 1957 report on patient H.M. — who lost the ability to form new conscious memories after bilateral medial temporal lobe surgery — follow-up experiments found he could still improve, day after day, at perceptual-motor tasks he had no recollection of ever practicing. Cohen and Squire made the dissociation exact in 1980: amnesic patients acquired a mirror-reading skill at the same rate as healthy controls and retained it for at least three months, while remaining impaired at remembering the words they had read — or, sometimes, the fact of having done the task at all. Their wording: amnesia spares knowledge “based on rules or procedures” while destroying knowledge that is “data-based or declarative.” Ryle’s distinction, honored by the nervous system.

So when I say a wiki-writing agent and a script-writing agent are externalizing different kinds of knowledge, I am not proposing a taxonomy. I am pointing out that the oldest taxonomy in memory research has quietly reappeared on our filesystems.

ACT-R, externalized onto a disk #

There is an even more direct ancestor. ACT-R, the cognitive architecture John Anderson’s group has developed since the 1980s, engineers a mind around precisely this split: a declarative memory of factual chunks and a procedural memory of production rules, with distinct dynamics for each. An agent that maintains a Markdown knowledge base alongside a library of scripts is, structurally, ACT-R with its two stores externalized onto a filesystem — chunks become pages, productions become programs. The pattern’s two poles are already visible in the wild. Karpathy’s LLM-wiki proposal is the declarative pole: the wiki as “a persistent, compounding artifact” of synthesized understanding. Voyager is the procedural pole: a Minecraft agent whose only long-term memory is an ever-growing skill library of executable code — skills that are “temporally extended, interpretable, and compositional” — which in its authors’ evaluation produced 3.3× more unique items and reached tech-tree milestones up to 15.3× faster than prior systems. And procedural artifacts seem to travel well: Memp finds that procedural memory distilled by a stronger model transfers to and improves a weaker one. The declarative counterpart exists too — Reflexion stores verbal self-critiques as text and reuses them — but notice that nobody runs Reflexion’s notes as a nightly job, and nobody asks Voyager’s skill library to explain itself.

Different metabolisms, not different encodings #

Why insist these are two minds rather than two serializations? Because the characteristic transformation each store rewards is different:

In words: the Markdown mind metabolizes what happens to it into claims that can be compared, contradicted, and merged; the tool mind metabolizes what it solves into operations that can be composed and re-executed. Each write biases the next one. A store full of essays makes the cheap next move “revise the synthesis”; a store full of scripts makes it “wrap two functions into a third.” Over a year, that compounding bias is a temperament.

The failure modes diverge the same way, which is how you can tell the systems apart from outside. Prose rots silently: a stale claim still reads fluently. Code rots loudly: a broken script throws — but its rationale evaporates, because nothing in fit_model.py records why that model and not another. One mind accumulates beliefs it can no longer justify executing; the other accumulates capabilities it can no longer justify believing. This is also why I think the choice of persistence format is not cosmetic: there is early evidence that the surrounding machinery shapes the store itself. A July 2026 study of filesystem memory reports that changing the tool set alone reshapes an agent’s memory store about as strongly as swapping the underlying model — a first observation from a single unreplicated preprint, but pointing exactly this direction: what an agent can do to its memory partly determines what kind of memory it grows.

Two hedges, both owed. First, the clean two-system story is a simplification even for brains: Squire himself has argued the field moved “beyond dichotomies” toward many interacting memory systems — striatal habits, priming, conditioning — not two. I expect the same of artificial minds; the two phenotypes here are the first cut of a larger design space I map in The Axes of Artificial Cognition. Second, Voyager lives in Minecraft, a world practically built out of procedures; how far the procedural advantage generalizes to messier domains is open. Whether the divergence between our two forked agents actually compounds — rather than washing out against the base model’s homogenizing prior — is the falsifiable bet I develop in Same Model, Different Minds. And why one architecture shouldn’t try to be both at once — why the objectives are antagonistic rather than merely different — is the subject of Tradeoffs Make Minds.

For now, the practical upshot is a hiring heuristic. When you choose what your agent persists — pages or programs — you are not choosing a file format. You are choosing which of two seventy-five-year-old kinds of knowledge it will get better at holding, and therefore what kind of colleague it will be a year from now.

What would change my mind: a controlled long-horizon study in which agents with essay-only and script-only stores converge to statistically indistinguishable behavior across explanation tasks and execution tasks alike — or evidence that a single mixed store matches both specialists at equal compute, which would demote the declarative–procedural split from an axis of mind design to a storage detail.