Solving the Clive Wearing Problem: One-Shot Episodic Memory for Frozen Transformers

Tommi Niemi / Rotko Networks

Hidden-state episodic memory for frozen transformers. No gradients.
Teach via one forward pass, recall via cosine similarity + logit injection.
200-line Python reproduction included.

pip install transformers torch numpy && python python/epimem.py

results/e2e_output.log

@@ -0,0 +1,16 @@
+=== Sleep Cycle Report ===
+NREM: 8 synapse updates
+  syn_input_attn: residual=0.0133
+  syn_attn_output: residual=0.0333
+  syn_basal_ganglia: residual=0.0605
+  syn_hippocampus: residual=0.0347
+  syn_insula: residual=0.0118
+  syn_motor_input: residual=0.0136
+  syn_cerebellum: residual=0.0161
+  syn_output_motor: residual=0.0446
+Graduated: 0 episodes → semantic
+REM: 3 replayed, 0 fears treated
+Health: 1.00/1.00
+=== Angeris Bounds Analysis ===
+Neurons: 64 total, dead: input=0 attn=0 output=0 motor=0
+  ALL SYNAPSES AT OPTIMUM — sleep consolidation won't improve further