🔔 CDM V6 HORN — Memory Slots That Ring

Competitive Docking Memory with Harmonic Oscillator Recurrent Nodes — DuoNeural, 2026

This 37M model uses 16 persistent memory slots per layer with second-order oscillator dynamics. Unlike standard transformers (attention-only) or simple RNNs (first-order decay), each slot is a damped harmonic oscillator — it can ring when struck by a salient input.

The model spontaneously discovers three slot regimes without any explicit supervision:

  • 🔔 Underdamped (Layer 0 + Layers 6-7): slots ring briefly/persistently — reactive and resonant
  • 🧲 Overdamped (Layers 1-5): slots absorb input smoothly and hold — stable long-range storage

Below: watch the Slot Logit Lens as the model generates, and see the Oscillator Panel showing what dynamics each layer learned.

37M params · val CE 1.5818 · new best at 37M on TinyStories

10 150
0.1 1.5
5 100
Try one of these

🧠 Slot Logit Lens — What Each Memory Slot Tracks

Each row = one generated token. Columns = 16 slots. Text = top-3 vocabulary tokens in that slot's hidden state (Layer 0). Watch slots spontaneously specialize.

⚙️ Learned Oscillator Dynamics (fixed after training)

Layer-by-layer breakdown of the per-slot damping (γ) and frequency (ω) parameters learned by gradient descent on language modeling. No explicit objective guided this — the model discovered that different temporal regimes serve different layers.

DuoNeural — open research lab · one human, two AIs, shared curiosity

Model: DuoNeural/CDM-V6-HORN-TinyStories-37M · Papers: zenodo.org/communities/duoneural · huggingface.co/DuoNeural