Contemporary large language models treat memory as passive storage—frozen in trained weights, transiently held in a context window, or externally indexed via retrieval-augmented generation. We argue this framing is incomplete. Drawing on neuroscience evidence that biological memory simultaneously performs world-modeling, predictive coding, and identity maintenance, we propose DMIA (Dynamic Memory-Integrated Architecture), a system in which memory is an active, predictive, relational substrate. D…
Read moreContemporary large language models treat memory as passive storage—frozen in trained weights, transiently held in a context window, or externally indexed via retrieval-augmented generation. We argue this framing is incomplete. Drawing on neuroscience evidence that biological memory simultaneously performs world-modeling, predictive coding, and identity maintenance, we propose DMIA (Dynamic Memory-Integrated Architecture), a system in which memory is an active, predictive, relational substrate. DMIA comprises four coupled components: (1) a Dynamic Semantic Graph (DSG) that continuously updates relational knowledge; (2) a Predictive Memory Loop (PML) that activates memory prior to perception; (3) a Hierarchical Temporal Consolidation (HTC) module inspired by hippocampal-cortical transfer; and (4) a Persistent Identity Core (PIC) that maintains coherent self-representation across sessions. We implement DMIA in Python and evaluate it on a synthetic continual-learning benchmark with 100 facts across 10 topics, comparing it against context-window, limited-RAG, and static-graph baselines. On an updated evaluation reporting Recall@5, NDCG@5, and Purity@5 across both learning phases, DMIA achieves perfect performance throughout (Recall@5 = NDCG@5 = Purity@5 = 1.000 on Phase 1, Phase 2 new topics, and Phase 2 old topics) with zero catastrophic forgetting (0.0 percentagepoint drop in both recall and NDCG). This compares to the context-window baseline, whose old-topic recall collapses from 92.0% to 8.0% (91.3% forgetting); limited RAG, whose old-topic recall collapses from 100.0% to 40.0% (60.0% forgetting); and the static graph, which—like DMIA—shows zero forgetting but at lower absolute recall (96.0% on both phases for old topics) and with a smaller memory footprint (67 vs. 84 nodes). The PIC exhibits near-zero drift (0.0005 per consolidation cycle), confirming stable identity maintenance. However, the PML’s predictive accuracy remains weak (11.1% next-topic prediction, barely above random), indicating that the forward model requires richer sequential structure or more sophisticated parameterization to realize its theoretical potential. We release the implementation and discuss scaling to real-world long-horizon tasks.