The Price of Meaning: Why Every Semantic Memory System Forgets

Organising memory by meaning makes forgetting and false recall inevitable. Scaling up does not fix it.

Sambartha Ray Barman, Andrey Starenky, Sophia Bodnar, Nikhil Narasimhan, Ashwin Gopinath (Sentra.app / MIT)

Overview

This repository contains the complete code, data, figures, and paper for the "No Escape" follow-up to the HIDE paper ("The Geometry of Forgetting"). We prove that any memory system satisfying the Semantic Proximity Property — semantically related items represented more similarly than unrelated items — must exhibit power-law forgetting, false recall, and partial retrieval states. We verify this across five architecturally distinct memory systems.

See docs/PROJECT_CONTEXT.md for the full motivation and docs/EXPERIMENT_LOG.md for the experimental record.

Repository Structure

├── noescape/                    # Core library
│   ├── architectures/           # 5 memory architecture implementations
│   ├── experiments/             # Ebbinghaus, DRM, Spacing, TOT, dimensionality
│   ├── analysis/                # Figure generation + statistics
│   ├── math/                    # Theorem verification
│   ├── solutions/               # Solution analysis (4 proposed cures)
│   └── utils.py                 # Shared utilities
├── results/                     # Raw experimental results (36 JSON files)
│   ├── vector_db/               # Architecture 1
│   ├── graph/                   # Architecture 4
│   ├── attention/               # Architecture 2
│   ├── parametric/              # Architecture 5
│   ├── filesystem/              # Architecture 3
│   ├── dimensionality/          # d_eff per architecture
│   ├── theorems/                # Theorem verification
│   ├── solutions/               # Solution analysis
│   └── verification.json        # Success criteria
├── figures/                     # All figures (PDF + PNG)
├── paper/                       # LaTeX source + compiled PDF
├── docs/                        # Project context + experiment log
├── run_calibration_v2.py        # Architecture 1 calibration
├── run_attention_full.py        # Architecture 2 full experiments
├── run_parametric_full.py       # Architecture 5 PopQA experiments
├── run_filesystem_full.py       # Architecture 3 experiments
├── run_remaining.py             # Supplementary experiments
├── config.yaml                  # All hyperparameters
└── requirements.txt             # Python dependencies

Quick Start

pip install -r requirements.txt

# Reproduce Architecture 1 (Vector DB) — ~1 hour
python run_calibration_v2.py

# Reproduce Architecture 2 (Attention) — ~4 hours
python run_attention_full.py

# Reproduce Architecture 5 (Parametric/PopQA) — ~30 min
python run_parametric_full.py

# Regenerate all figures from raw data
python -c "from noescape.analysis.figures import generate_all_figures; generate_all_figures('results', 'figures')"

Requirements

• NVIDIA A100 80GB GPU (or equivalent)
• Python 3.11+
• ~10 GPU-hours for full reproduction

Models Used (all open-weight)

• BAAI/bge-large-en-v1.5 (MIT) — 1024-dim sentence embeddings
• sentence-transformers/all-MiniLM-L6-v2 (Apache 2.0) — 384-dim sentence embeddings
• Qwen/Qwen2.5-7B-Instruct (Apache 2.0) — LLM for attention/parametric/filesystem

Key Results

Architecture	Forgetting b	DRM FA	d_eff
Vector DB	0.440 ± 0.030	0.583	158 / 10.6
Graph	0.478 ± 0.028	0.208	127
Attention	phase transition	0.000†	17.9
Parametric	0.215*	0.000†	17.9
Filesystem	0.000	0.000	158

* PopQA interference b. † Behavioural; geometric prediction holds (24/24 caps).

Citation

@article{barman2025priceofmeaning,
  author = {Barman, Sambartha Ray and Starenky, Andrey and Bodnar, Sophia and Narasimhan, Nikhil and Gopinath, Ashwin},
  title = {The Price of Meaning: Why Every Semantic Memory System Forgets},
  year = {2025}
}

Acknowledgements

Computational experiments and manuscript preparation were assisted by Claude (Anthropic).