Recognition Without Recall: Empirical Limits of Biologically-Inspired Memory Consolidation in Transformers

We implement and empirically evaluate a biologically-faithful memory consolidation architecture forLarge Language Models, drawing on the synaptic tagging-and-capture hypothesis Frey and Morris, 1997,Redondo and Morris, 2011, the Complementary Learning Systems framework McClelland et al., 1995,and hippocampal-neocortical replay during sleep Wilson and McNaughton, 1994, Diekelmann and Born,2010. Our system, SLEEP, augments a 7B-parameter transformer with five interacting components:prediction-error tagging, direct-write key-value (KV) memory, protein-resource-allocation prioritization,generative replay, and LoRA-based consolidation under explicit safety constraints. Across 200 singleexposure synthetic facts on Qwen2.5-7B, we systematically isolate where the architecture succeeds andfails. Direct-write KV memory injection produces a clear recognition signal (Tagged–Untagged ∆ = +0.16on multiple-choice), but the same memories do not surface during free-form autoregressive recall—atransformer analogue of the recognition–recall dissociation in cognitive psychology. Even when generativereplay is replaced with the original tokens (eliminating replay-quality as a confound), SLEEP’s safetymachinery yields no operating point with BCP 0.05. In a multi-cyclecontinual-learning evaluation on disjoint batches, however, SLEEP preserves base capability 2× betterthan naive LoRA fine-tuning at every cycle (BCP 2.31 vs 4.74 at cycle 3), confirming that single-cyclebenchmarks understate biologically-inspired safety mechanisms. We map five formalization amendmentsrequired for the architecture to function at scale and identify the consolidation step’s stability–plasticitytradeoff as the central open problem.