The Geometry of Forgetting

Code and data for: The Geometry of Forgetting — showing that high-dimensional embedding spaces, subjected to noise, interference, and temporal degradation, reproduce quantitative signatures of human memory with no phenomenon-specific engineering.

Authors: Sambartha Ray Barman, Andrey Starenky, Sophia Bodnar, Nikhil Narasimhan, Ashwin Gopinath

Quick Start

# Install dependencies
pip install -r requirements.txt

# Run all experiments (requires 4x A100 GPUs)
python scripts/run_all.py

# Generate all figures from cached results
python scripts/generate_figures.py

# Compile paper
cd paper && pdflatex hide_paper.tex && cd ..

Repository Structure

submission/
  paper/                    # Manuscript and figures
    hide_paper.tex          # Main LaTeX source
    hide_paper.pdf          # Compiled PDF
    figures/                # All figures (6 main + 11 extended data)
  hide/                     # Core library
    core/                   # HIDESpace, temporal encoding, interference, consolidation
    models/                 # Embedding models, Qwen adapter
    utils/                  # Metrics, data loading, visualization
  experiments/              # Experiment scripts (one directory per phase)
    phase1/                 # bAbI reasoning (memory retrieval)
    phase2/                 # Temporal memory (Ebbinghaus forgetting)
    phase3/                 # Consolidation and interference
    phase4/                 # Cross-modal binding
    phase5/                 # Emergent phenomena (DRM, spacing, topology)
    spectral/               # Effective dimensionality analysis
  results/                  # All experimental results (JSON + CSV)
    phase1/ ... phase5/     # Per-phase results with 5-seed replication
    spectral/               # Dimensionality and MiniLM interference results
    interference/           # Interference theory experiment results
    spacing_sweep/          # Spacing effect sweep results
    topology/               # Persistent homology results
  configs/                  # YAML hyperparameter configs per phase
  scripts/                  # Reproduction scripts
    run_all.py              # Master experiment runner
    generate_figures.py     # Figure generation from results
    figure_style.py         # Nature-quality matplotlib settings
    gen_fig*.py             # Individual figure generators
  data/                     # DRM word lists + download script
  tests/                    # Sanity tests
  LICENSE                   # Apache 2.0
  requirements.txt          # Pinned dependencies

Hardware Requirements

• Full reproduction: 4x NVIDIA A100 (80GB) GPUs
  • GPU 0: Qwen2.5-7B (answer generation)
  • GPU 1: Embedding models (MiniLM / BGE-base / BGE-large)
  • GPU 2: CLIP + parallel experiments
  • GPU 3: FAISS-GPU + batch compute
• Figures only: CPU (regenerate from cached results in results/)
• Tests only: CPU

Models (All Open-Weight)

Model	HuggingFace ID	Use	License
Qwen2.5-7B	Qwen/Qwen2.5-7B	Answer generation	Apache 2.0
MiniLM-L6-v2	sentence-transformers/all-MiniLM-L6-v2	Text embedding	Apache 2.0
BGE-base	BAAI/bge-base-en-v1.5	Text embedding	MIT
BGE-large	BAAI/bge-large-en-v1.5	Text embedding (scale)	MIT
CLIP ViT-B/32	openai/clip-vit-base-patch32	Image embedding	MIT

Datasets (All Public)

Dataset	Source	License
bAbI QA (en-10k)	Muennighoff/babi	BSD
TempLAMA	Yova/templama	MIT
CIFAR-100	torchvision.datasets.CIFAR100	BSD
COCO Captions	jxie/coco_captions	CC BY 4.0
Flickr30k	lmms-lab/flickr30k	Research
Wikipedia (en)	wikipedia (20220301.en, streaming)	CC BY-SA
DRM word lists	Roediger & McDermott (1995)	Public domain

Reproduction

Full pipeline (4x A100)

# Download datasets
bash data/download_data.sh

# Run all experiments across 5 seeds [42, 123, 456, 789, 1024]
python scripts/run_all.py

# Generate figures
python scripts/generate_figures.py

Figures only (CPU)

# Results are pre-cached in results/
python scripts/generate_figures.py

Run tests

python -m pytest tests/ -v

Key Results

Phenomenon	Observed	Human	Notes
Forgetting exponent	0.460 +/- 0.183	~0.5	Interference-driven, not decay
DRM false alarm rate	0.583	~0.55	Unbaked — zero parameter tuning
Spacing effect	massed < short < med < long	Same ordering	Boundary-condition-dependent
Effective dimensionality	d_eff ~ 16	d = 100-500 (cortex)	All models concentrate in ~16 dims
Tip-of-tongue rate	3.66%	~1.5%	Qualitative emergence

Seeds and Reproducibility

All experiments use seeds [42, 123, 456, 789, 1024] with bootstrap 95% CIs (10,000 resamples). Results are deterministic given the same seed, model weights, and hardware.

Citation

@article{geometryofforgetting2025,
  title={The Geometry of Forgetting},
  author={Ray Barman, Sambartha and Starenky, Andrey and Bodnar, Sophia and Narasimhan, Nikhil and Gopinath, Ashwin},
  year={2025},
  note={Under review}
}

License

Apache 2.0. See LICENSE.

Acknowledgements

Code generation assisted by Claude (Anthropic).