A Reproducible, Dual-Readout Hippocampal–Cortical Complementary Memory Model (HCM): An Empirical Log-N Collapse Regime Map
We present HCM (Hippocampal–Cortical Complementary Memory), a biologically inspired computational model with an executable, seed-reproducible reference implementation. HCM factorises memory into nine operators: encoding, episodic and semantic stores, working memory, consolidation, retrieval control, forgetting, and gating. Three claims. (i) A log-N collapse regime. The harm of retrieval inverse temperature β grows with library size as log N; two earlier negative results were artifacts of one mis-set β; the staircase is measured 10³–10⁶, the 10⁶ tier passing a gate registered before its data exist (60/60 both levels, both arms, zero abstention); an online β* transfers across sessions at 99.6% of hand-tuned performance; a six-architecture contrast passes its three pre-registered gates — classical attractors collapse to the floor, modern baselines sit flat, only HCM's log-N β*(N) compensation holds retrieval near 0.8–0.9. The n=4 staircase fit is descriptive (r²=0.800); the informative prediction — β* shifting with code overlap — remains unmeasurable on the real corpus (G-R3 FAIL). (ii) Read each projection twice. A k-sparse code for writing, the dense evidence as a Hopfield attractor's initial state: at 60% dropout sparse readout degrades (50/50→39/50), dense holds; the gate abstains from novel cues (d′=7.12); at 92,760 items the semantic channel carries no identity (0/200) while attractor control stays healthy. (iii) Consolidation is valued by interleaving, not prioritisation. Without replay, old-task accuracy collapses (0.0/24, 20 seeds); interleaved retains 13.85/24; prioritised 14.80/24, no detectable difference — a negative result. On real colloquial Chinese, 2,000-item stores reach the synthetic ceiling only under a corrected judging criterion; at 92,760 items recall is 129/200 at 30% degradation. A controlled synthetic follow-up confirms the direction of the informative prediction (mean β* rising 20→200 across overlap 0→0.8 at N=2×10³) but fails its own pre-registered monotonicity bar (pooled Spearman ρ=0.838<0.9), saturating at low overlap (§4.21). All experiments run on a NumPy-only implementation with fixed seeds.
Authors
- Haowei Wang (ORCID: https://orcid.org/0009-0004-2241-5511)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24359580
- Primary Topic
- Memory and Neural Mechanisms
- Type
- preprint