We present a memory-augmented transformer in which attention serves simultaneously as a retrieval, consolidation, and write-back operator. The core update, A⊤AVW, re-grounds retrieved values into persistent memory slots via the Gram matrix A⊤A, providing a principled tripartite projection: observation space → latent memory → supervised transformation. We partition the memory into lateralized left and right banks coupled through a sign-controlled cross-talk matrix Ws, and show that the sign of this coupling is decisive for specialization. Excitatory cross-talk (s=+1) causes bank-dominance collapse: one bank monopolises all inputs and Pct→0.5, despite lowering task loss. Inhibitory cross-talk (s=−1), motivated by the net inhibitory effect of callosal projections in human cortex, actively suppresses contralateral bank activation and achieves saturated specialization (Dsep=±1.00, Pct≈0). On a controlled symbolic benchmark combining an episodic bijection cipher (requiring associative recall) with a strict arithmetic progression (requiring rule extraction), the inhibitory model reduces cipher-domain loss by 124× over the baseline while matching it on the arithmetic domain, confirming that persistent lateralized memory is necessary for episodic recall but not for rule-based prediction.
Figures & tables
Figure 1: The tripartite projection sequence mapping latent values to the observation space and back, before applying the supervised transformation.
Figure 2: Bidirectional cross-talk pathways for lateralized memory (inhibitory mode, s=−1 , proposed). The dashed red lines show the paths where one bank’s values reach the contralateral update. Under inhibitory cross-talk, these paths carry a negative coefficient: the right bank’s values Vr suppress the left bank’s update (and vice versa), sharpening bank separation rather than sharing representations. Setting s=+1 reverses these signs to excitatory; setting the cross-weights to zero gives the split-brain baseline.
Model
Accuracy
Loss
Datasetl
Datasetr
Datasetlr
Datasetl
Datasetr
Datasetlr
Standard Transformer
100.0%
100.0%
93.8%
0.0747
0.0002
0.1692
Attention-Coupled Lateral (Ours)
100.0%
100.0%
94.4%
0.0006
0.0002
0.1452
Table 1: Next-token prediction accuracy and cross-entropy loss on homogeneous and interleaved symbolic sequences. The lateralized memory prevents the performance collapse seen in the baseline on mixed tasks.
Figure 3: Training convergence of the Attention-Coupled Lateral model over 50 epochs. Panel 1 (task loss) : cross-entropy drops from 1.12 to 0.03. Panel 2 (total loss) : routing auxiliary term drives total loss slightly negative once routing saturates. Panel 3 ( Dsep ) : Separation Degree converges after brief instabilities at epochs 15 and 33. Panel 4 ( Pct ) : Cross-Talk Penalty collapses to ≈0 within 4 epochs and stays there, confirming stable lateralized routing throughout training.
Metric
Datasetl
Datasetr
Datasetlr
Left Separation Degree Dsep(L)
+ 1.00
− 1.00
− 0.42
Right Separation Degree Dsep(R)
− 1.00
+ 1.00
+ 0.42
Cross-Talk Penalty Pct
0.00
0.00
0.03
Table 2: Lateralization metrics demonstrating the emergence of specialized memory banks. Higher Dsep indicates strong routing to the correct hemisphere, while low Pct indicates minimal inappropriate attention cross-talk.
Cross-talk Mode
Dsep(L)∣l↑
Dsep(L)∣r↓
Pct↓
Acc lr↑
Loss lr↓
None (split-brain)
+ 1.00
− 1.00
0.00
94.4%
0.1456
Excitatory ( s=+1 )
− 0.82
− 0.93
0.46
95.8%
0.1017
Inhibitory ( s=−1 , ours)
+ 1.00
− 1.00
0.03
94.4%
0.1454
Table 3: Cross-talk mode ablation on the mixed dataset Datasetlr , evaluated with Dsep(L) (measured on Datasetl and Datasetr separately), Pct (on Datasetlr ), and accuracy/loss on Datasetlr . All conditions share the same architecture, training procedure, and routing auxiliary loss.
We present a miniature brain transformer architecture that extends the attention-coupled latent memory framework with four additional brain-region analogues: a thalamic relay, an amygdaloid salience module, a prefrontal working-memory (PFC) buffer, and a cerebellar fast-path, all coupled by inhibitory callosal cross-talk between lateralized hippocampal banks. We evaluate on a two-domain benchmark -- MQAR (Multi-Query Associative Recall; episodic domain) and modular arithmetic (+1 mod 10; rule-based domain) -- using a seven-variant additive ablation. The central empirical finding is a surprise: inhibitory callosal coupling alone never lateralizes the banks (variants 1-5 maintain D_sep ~ 0.25 and P_ct ~ 0.25 for all 30 epochs). Functional lateralization requires the synergy of PFC and inhibition: only when the PFC buffer is added (variant 6) does a sharp, discontinuous phase transition fire -- at epoch 11 for the PFC-only variant and epoch 10 for the full model -- collapsing P_ct from 0.25 to ~0.002 and more than doubling D_sep from 0.251 to 0.501 in a single gradient step. The PFC buffer acts as a symmetry-breaker: its slowly drifting domain context creates the initial asymmetry that the inhibitory feedback loop then amplifies irreversibly. The cerebellar fast-path accelerates the transition by one epoch (epoch 10 vs. epoch 11) with no asymptotic change, confirming its convergence-acceleration role. The result constitutes a novel, falsifiable prediction -- no lateralization without working memory context -- and a principled, neurobiologically motivated blueprint for hierarchical persistent memory in sequence models.
Hong Jeong
Department of Computer Information Engineering, Inha University in Tashkent, Uzbekistan
Copying short-term memory (STM) into a slower store can preserve state across a context boundary, but persistence alone does not ensure that the retained state influences subsequent memory access. We test this distinction in a Phasor Memory Network (PMNet) using Consolidator, a shared slot-local operator that transforms routed STM before accumulating it into long-term memory (LTM), without replaying the source tokens. After each consolidation, the KV cache and STM are cleared. The retained LTM can still be read and is also fed into the hierarchical router, thereby conditioning which explicit-memory slots subsequent inputs access. We evaluate this mechanism on a two-segment modulo-10 mapping task in which the second segment updates the mapping at the same memory address. Following a second consolidation and reset, a held-out query must recover the updated mapping from LTM. The backbone and memory interface are frozen, leaving only 12.35K Consolidator parameters trainable (0.041% of a 29.95M model). Across five paired runs from the same STM-pretraining checkpoint, direct LTM routing raises updated-mapping recall from 44.38±1.94% to 87.02±1.76% (+42.64±1.10 percentage points), while immediate STM recall remains 89.90% in both conditions; both train separate Consolidators and retain the same LTM read paths. Learned consolidation outperforms forced identity accumulation by 21.40±1.91 percentage points without routing and 68.70±1.76 with routing. Thus, on this task, consolidated LTM serves as both retrievable content and an access state that shapes subsequent slot selection.
Sungwoo Goo, Hwi-yeol Yun, Sangkeun Jung
College of Pharmacy, Chungnam National University · Department of Computer Science & Engineering, Chungnam National University
Linear attention replaces the growing key--value (KV) cache of standard attention with a fixed-size recurrent state, substantially reducing memory growth with context length. This efficiency, however, changes how past information is stored: many tokens must share and repeatedly update the same memory. We study how this recurrent state functions as a memory system. Using an analytical decomposition together with controlled causal interventions in pretrained GLA and GDN models, we trace how recalled information is written, retained, and later accessed. We find that fact-specific information enters recurrent memory through concentrated, content-dependent writes and is later accessed through concentrated query-time read pathways. Multiple facts can remain selectively accessible within the same state, yet their internal representations exhibit cross-fact causal coupling rather than independent KV-like storage. As memory load increases, both recall and targeted editability degrade, whereas elapsed context alone has a substantially smaller effect within the tested regime. Causal interventions on subsequent writes further show that interference is shaped by their overlap with existing memory. Finally, in hybrid architectures that combine recurrent layers with full attention, the runtime memory directly supporting recall shifts predominantly to the full-attention KV state. Together, these results reveal how fixed-size recurrent memory supports selective recall despite shared storage, while exposing the interference and capacity limits that distinguish it from token-addressable KV memory.