MCQ answer: (1) — improved daily, remembered none of it
H.M.’s tracing speed and accuracy improved measurably day after day — a completely normal procedural learning curve. Yet each day, he reported no memory of ever having done the task before, often expressing surprise at how well he performed on something he insisted was novel. This is the single most famous demonstration in memory research that skill learning (procedural memory) and event memory (episodic memory) are dissociable — genuinely separate systems, not one memory faculty operating at different strengths. Option (0) describes what you’d predict if memory really were one unified system that his damage weakened globally; it’s exactly what didn’t happen. Option (2) describes ordinary, non-dissociated learning — not what was observed. Option (3) is simply false — mirror-tracing is a well-established procedural learning task specifically because performance improves with practice via a memory mechanism, just not a declarative one.
Part 2 — Why the pattern is the strong evidence
A uniformly mild deficit across all memory types is exactly what you’d expect from “memory is a single resource, and this damage weakened it somewhat” — consistent with one system, just running worse. What H.M. showed instead is a double-dissociable pattern: one capacity (procedural learning) performing at essentially full, healthy strength, while another (new episodic memory) performed at essentially zero — with no continuum of “medium impairment” in between. That specific combination — one system spared, the other devastated, by the same focal brain damage — is much harder to explain under a single-unified-memory account than under an account where procedural and declarative memory depend on different, at least partially separable, neural substrates (the hippocampus being critical for forming new declarative memories, but not required for procedural learning, which appears to route through different structures, notably the basal ganglia and cerebellum). A single damaged region taking out one whole system while leaving another fully intact is close to the strongest kind of evidence neuropsychology can offer for separateness.
Part 3 — A predicted second dissociation: episodic vs. semantic
Yes — this dissociation has in fact been observed in other real cases (distinct from H.M. specifically). A plausible real-world pattern: a patient with more selective, milder hippocampal damage (as opposed to H.M.’s extensive removal) who cannot reliably form new episodic memories — they can’t tell you what they did yesterday, or recall the specific occasion they learned something — but who can still slowly acquire new semantic facts over repeated exposure, ending up “knowing” a new fact (e.g., a new person’s job title, repeated across many encounters) without being able to recall any specific instance of learning it. This is source amnesia taken to its logical extreme: retaining the fact while permanently losing the episodic “when/where I learned this” tag. (The developmental case of patients with early, selective hippocampal damage acquiring largely normal semantic/school knowledge despite profound episodic memory deficits — reported by researchers including Faraneh Vargha-Khadem — is a real-world instance of close to this pattern.) This fits the general logic of Part 2: if procedural/declarative can dissociate via H.M.’s damage, and episodic/semantic are meaningfully distinct systems rather than the same thing at different grain-sizes, a different lesion profile predicts they should be separately impairable too — and evidence broadly bears that prediction out, though less starkly clean than the H.M. case.
Part 4 — Dissociation in an agent harness
Take the episodic (context window) vs. semantic (trained-in weights) pair. A concrete dissociation: end a conversation (the context window is cleared) — everything episodic about this specific interaction (what the user just said, what was tried and failed, the specific back-and- forth) is gone, completely and immediately, the harness equivalent of H.M.’s inability to retain new episodic content. Yet the model’s general knowledge — facts, language ability, learned behavioral patterns baked in during training — persists completely unaffected across that same reset, the harness equivalent of intact semantic and procedural memory. This is in fact the default, unremarkable behavior of most LLM systems without a persistent memory feature — every new conversation starts in a state analogous to H.M.’s: full general competence, zero memory of prior specific episodes. Systems that add a persistent “memory” feature (like this very memory system being used to author these lessons) are explicitly engineering a workaround for exactly this default dissociation — building an artificial episodic store precisely because the base system, by default, has none.
The pattern
| System | What it stores | H.M.’s status | ACT-R analogue |
|---|---|---|---|
| Episodic | specific events (what/when/where) | severely impaired (new) | declarative chunks |
| Semantic | general facts/concepts | impaired for new learning | declarative chunks |
| Procedural | skills, how-to | intact, normal learning curve | production rules |
Rule: long-term memory is not one system with a single dial — episodic, semantic, and procedural memory are dissociable, meaning focal damage can devastate one while leaving another fully intact. The declarative/procedural split maps directly onto ACT-R’s chunks-vs-productions architecture from Stage 2, giving this psychological finding a concrete computational counterpart.
Where this goes: Stage 3 closes here, having covered attention’s gate (Lesson 16) and what happens once things get through it into durable storage. Stage 4 turns to what happens with heavy, repeated use of stored knowledge — the power law of practice, and how expertise itself is built.