A Log-N Collapse Law in Attractor Memory: A Reproducible, Dual-Readout Hippocampal–Cortical Complementary Memory Model (HCM)
Year: 2026
Type: preprint
Abstract: We present HCM (Hippocampal–Cortical Complementary Memory), a biologically inspired computational model of memory with an executable, seed-reproducible NumPy reference implementation and a systematic benchmark suite. HCM factorises memory into nine operators: input/context space, encoder, episodic fast store, semantic slow store, working memory, consolidation, retrieval control, forgetting, and neuromodulatory gating. Three claims are central. (i) A collapse law: the harm of the retrieval inverse temperature β grows with library size as log N; two earlier negative results were artifacts of a single mis-set β, disappearing at β=100 (dose–response, analytic theory, external replication), with the β≈20 turning point replicating across two encoders on a real corpus. (ii) A single projection should be read out twice: a k-sparse code for writing, dense evidence as the initial state of a modern Hopfield attractor for retrieval — at 60% dropout, sparse readout degrades (50/50→39/50) while dense readout stays 50/50. (iii) The value of systems consolidation lies in interleaving rather than prioritisation: without replay, old-task accuracy collapses to 0.0/24; interleaved replay retains 13.85±2.66 of 24; prioritised replay scores 14.80±3.96 — no detectable difference, a negative result we report. End-to-end on real colloquial Chinese, 2,000-item stores reach the synthetic ceiling (200/200 at 30% degradation) only under a corrected judging criterion, while the same model at 92,760 stored items retrieves 129/200. All experiments run on a NumPy-only reference implementation with fixed seeds; four machine-checkable acceptance gates are all green (numeric traceability 444/444, citation verification 447/447, closure 56/56, checksums 677/677).
Author: Wang, Haowei
Language: English
Topic: Memory and Neural Mechanisms
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7236761145
