L5 — dual learning systems: a fast episodic store and a slow semantic store
Two stores separate by persistence, not abstraction. Offline replay consolidates one-shot episodes into a separate, persistent slow store that generalizes to novel instances and greatly reduces catastrophic forgetting (dual ≪ single at every load). Supplied as an independent context channel, that slow store closes L4’s one open limit: strict recovery returns to ~oracle (≈ 1.0) where a content-derived gate degrades (0.68).
The wave machine builds the brain’s complementary systems — a fast episodic field (one-shot, high plasticity) that replays into a slow semantic field (low plasticity) during offline settling. The headline is C3: as the independent context channel L4 named, the slow store escapes the derivability⊥separability trade-off exactly, robust to context-cue corruption ≤ 0.3. The honest [O]: the textbook double dissociation does not hold — the fast Hebbian store already generalizes for free.
L5 is the direct sequel to L4’s single open limit. The brain pairs a fast hippocampal store (episodic, one-shot) with a slow cortical store (semantic, generalizing), consolidated by replay during sleep. The wave machine builds both, and the slow store turns out to be precisely the context channel L4 required (session v0.6, digest e8a623…).
Four results: three verified, one honest open
| result | finding | grade |
|---|---|---|
| C1 consolidation | replay builds a separate, persistent slow store that generalizes to novel instances (overlap > single-instance baseline 1−2ρ) | [V] |
| C2 forgetting | dual system reduces catastrophic forgetting vs a single leaky store — dual ≪ single at every load (gap > 0.2) | [V] |
| C3 context channel (headline) | as an independent context channel the slow store restores strict recovery to ~oracle: strict ≈ 1.0 ≫ content ≈ 0.68 ≫ flat 0; robust to ctx-corruption ≤ 0.3 | [V] |
| C4 double dissociation | does not hold — the fast store already generalizes for free | [O] |
How it closes L4
That open limit was structural: a gate derived from content alone cannot have both derivability and strict capacity. The slow semantic store breaks the trade-off because it is learned slowly over many instances and therefore supplies a context signal not derived from the current cue’s content. Fed in as that independent channel, strict instance recovery returns to oracle level exactly where content-derivation degraded — the L4 [O] becomes [V], as L4 itself predicted.