---
title: Memory Tape Pocket Lab
emoji: 🧪
colorFrom: indigo
colorTo: blue
sdk: static
app_file: index.html
pinned: false
---

# Memory Tape Pocket

Memory Tape Pocket is a compact differentiable-memory retest inspired by the
content-addressing mechanism of Neural Turing Machines. It learns random
key-value associative recall on tapes containing two to eight slots, then faces
unseen tapes with 16 and 32 slots.

The control is a larger fixed-state GRU trained on the same batches. The
interactive Space exposes the complete external tape and the learned read
weight assigned to every slot.

## Verified result

Across three independent training seeds, the 4,673-parameter content-addressed
model achieved **100% exact recall** on 8-, 16-, and 32-slot tapes. At 32 slots,
four times the maximum training length, its read head placed **99.974%** of its
attention on the correct slot.

The larger 5,584-parameter fixed-state GRU reached 13.51% accuracy at eight
slots, 7.66% at 16 slots, and **4.60% at 32 slots**. This benchmark isolates the
inductive bias of external content addressing; it does not claim the tiny model
implements every component of a full Neural Turing Machine.

```bash
uv run python projects/memory-tape-pocket/train.py
uv run pytest tests/test_memory_tape_pocket.py
```

## Hosted showcase

This free static Space preserves the complete original Gradio source, trained artifacts, evaluation files, and local launch requirements. Hugging Face now requires PRO for CPU-backed Gradio hosting, so the public landing page is static while the checked-in `app.py` remains the authoritative runnable demo.