Memoir AI Study: Writing to Memory While Thinking Slows Learning by 20%

Researchers Jaber and Jaber introduce Memoir, a method that combines per-sample fast memory, shared slow parameters, variable-depth latent recurrence, and a future-latent energy objective. The central question: should a model be allowed to write to its fast memory during the same pondering iteration that reads it? The team tested this 'coupled' approach against a read-only baseline using identical architectures — both with 81,738 parameters (76,362 trainable) and matched compute budgets. On a procedural associative recall task with key interference, the coupled arm achieved recall of 0.5203 (95% CI: [0.4522, 0.5883]) after 240 training steps across 12 seeds, while the read-only arm reached 0.6557 (CI: [0.5953, 0.7160]). The paired difference of 0.1354 gave a t-statistic of 3.23 (11 df, p<0.01), with the read-only version winning on 10 of 12 seeds.

However, after 960 training steps both arms hit 1.000 recall, indicating the penalty is in learning speed, not ultimate capability. The predicted failure where memory rewriting corrupts the energy signal did not occur — the energy margin actually grew and held. Additionally, kernel restructuring reduced the delta-rule forward pass time from 0.907 ms to 0.351 ms on the test device, a 2.6x improvement. These results suggest that models can eventually overcome self-interference from writing, but architects should consider deferring memory updates until after pondering to accelerate early training. The full paper includes 9 pages, 4 figures, and a link to code.