DRM substrate

Reproduction code for the manuscript The architecture of false memory: a non-parametric account of DRM and the FTT–AM dissociation, currently under double-blind review at Open Mind: Discoveries in Cognitive Science.

The substrate is a non-parametric memory architecture (sparse-binary vector-symbolic representations, content-addressable episodic and semantic stores, replay-driven consolidation, dual-path retrieval) that reproduces canonical DRM false-memory phenomenology without any learned parameters. All experiments reported in the manuscript can be regenerated from this repository.

Layout

cgm/                          # the memory substrate (DRM-relevant subset)
    vsa/                      # sparse-binary hypervectors + operations
    memory/                   # content-addressable stores + HNSW index
drm_paper/                    # paper experiments
    stimuli.py                # 55-list Roediger 2001 / Stadler 1999 battery
    encoding.py               # four encoders: random, Brown, ConceptNet, Nelson
    retrieval.py              # iterated retrieval + Remember/Know judgment
    experiment.py             # single-seed DRM experiment runner
    battery.py                # multi-seed battery + bootstrap CIs
    figures.py                # paper figure generators
    sweeps.py                 # fidelity-generalization 2-d sweep
    analysis.py               # recognition-rate summary + R/K dissociation
    run_headline_battery.py   # ConceptNet 55-list battery (fig2, fig3c, fig3d)
    run_nelson_battery.py     # Nelson encoder 55-list battery
    run_evidence_sweep.py     # multi-N_max sweep (fig4)
    run_keyvalue_experiment.py # pure key-value dissociation (fig_keyvalue)
    data/
        build_stadler_data.py        # transcribes Roediger 2001 Appendices A+B
        build_mediated_lures.py      # auto-generates mediated lures via ConceptNet
        build_nelson_norms.py        # parses USF free-association norms
        stadler1999_55lists.json     # 55 DRM lists with per-list BAS values
        mediated_lures_55.json       # one mediated lure per list
    tests/                    # pytest suite (109 tests)
    results/figures/          # the 9 figures included in the manuscript

Setup

Python 3.10+. The substrate has no GPU dependency.

python -m venv .venv
source .venv/bin/activate
pip install -e .

The four external dependencies are numpy, matplotlib, nltk, and (for tests) pytest. They install automatically from pyproject.toml.

External data

Two data files are required for the headline experiments but are not redistributed in this repository:

1. ConceptNet Numberbatch (English, 19.08 release, ~311 MB). Used as the primary perceptual encoder. Download:

   curl -L -o drm_paper/data/numberbatch-en-19.08.txt.gz \
     https://conceptnet.s3.amazonaws.com/downloads/2019/numberbatch/numberbatch-en-19.08.txt.gz
   

2. Nelson free-association norms (USF norms, redistributed at github.com/teonbrooks/free_association). Used by the Nelson-encoder ablation:

   curl -sL https://github.com/teonbrooks/free_association/archive/refs/heads/master.zip \
     -o /tmp/fa.zip
   unzip -q /tmp/fa.zip -d /tmp
   cp /tmp/free_association-master/free/free_association.txt \
      drm_paper/data/nelson_norms_raw.txt
   python drm_paper/data/build_nelson_norms.py
   

The DRM lists (stadler1999_55lists.json) and mediated lures (mediated_lures_55.json) ship with the repository.

Reproducing the paper figures

Each headline figure has a single entry-point script.

Figure	Command
fig2_recognition_rates.{pdf,png} (ConceptNet, 20 canonical lists)	see one-liner in §5.1 of the paper, or run figure_recognition_rates on a 20-list ConceptNet battery
fig2_recognition_rates_nelson.{pdf,png}	python drm_paper/run_nelson_battery.py --n-seeds 20
fig3_per_list_fa.{pdf,png}	bundled in run_headline_battery.py
fig3c_bas_regression_evidence.{pdf,png}	python drm_paper/run_headline_battery.py --n-seeds 20
fig3c_bas_regression_nelson.{pdf,png}	python drm_paper/run_nelson_battery.py --n-seeds 20
fig3d_human_evidence.{pdf,png}	python drm_paper/run_headline_battery.py --n-seeds 20
fig4_evidence_by_iteration.{pdf,png}	python drm_paper/run_evidence_sweep.py --n-seeds 10
fig4b_ftt_am_dissociation.{pdf,png}	uses the same multi-N_max sweep
fig6_frontier_lure.{pdf,png}	python drm_paper/sweeps.py
fig_keyvalue_dissociation.{pdf,png}	python drm_paper/run_keyvalue_experiment.py --n-seeds 20

All runners cache the underlying batteries as *.pickle files in drm_paper/results/, so re-running with --load-cached regenerates figures in seconds.

Tests

pytest drm_paper/

All 109 tests run in roughly 5 minutes on a laptop (the slowest individual test builds the Brown-corpus co-occurrence matrix on first run; subsequent runs are cached by NLTK).

License

MIT. See LICENSE.

Author

Anonymized for double-blind peer review. Will be deanonymized on acceptance.