User as Engram: Internalizing Per-User Memory as Local Parametric Edits

Bojie Li · Pine AI

 Paper · Interactive site

A user's facts become a few rows in a content-addressed memory table — not a rewrite of the model. Personal memory splits into content (the user's specific facts) and reasoning skill (turning facts into answers); we store each the right way. Per-user content is written as surgical edits to the hash-keyed memory table of an Engram model, and the reasoning skill lives in one shared adapter that everyone uses — an artificial analogue of the brain's complementary learning systems.

Headline results (vs. a per-user LoRA, on Mini-Engram-d20):

• ≈5.6× (up to 7.4×) higher indirect-reasoning accuracy at matched direct recall.
• ≈33,000× less disruption to unrelated text — the edit is local by design (exactly 0.000 off-trigger), not merely small.
• 88 KB per user instead of 14.2 MB, with the backbone left bit-identical and zero cross-user leakage by construction.
• Against retrieval, which method wins is set by deployment: a per-user table doesn't grow with the population, so past ~100 facts/user it overtakes a retrieval pipeline on a 2.5× larger model.

Repository layout

Path	Contents
paper/	LaTeX source (main.tex), figures, and figure-generation scripts.
code/	Everything to run the experiments: the training/eval harness (nanochat_harness/, vendored, with our Engram module + GPT edits), our experiment scripts (scripts/), and the per-user-LoRA baseline + data generators (lora_baseline/). See code/README.md.
data/	Synthetic per-user fact sets (users/, users_medical/) and fact corpora — all included. See data/README.md.
results/	Result JSONs the paper figures are built from. See results/README.md.
site/	The interactive site (React + Vite). See site/README.md.
docs/dev-notes/	Working research logs and historical scripts (provenance; not needed to use the method).

The repo is self-contained: the harness, code, and synthetic data are all here, so reproduction needs no external checkouts. Only the trained Mini-Engram checkpoints (178 M – 1.22 B) are too large to ship — regenerate them by training Engram with code/ (see REPRODUCE.md).

Reproducing

Full guide: REPRODUCE.md (four levels, from laptop to full training run). The fastest path rebuilds every paper figure from the committed results/, no GPU:

cd paper
pip install -r requirements.txt
bash make_all_figures.sh        # 11 scripts → figs/*.pdf (all 38 figures)

The synthetic data (data/), all result JSONs (results/), and the full experiment + system code (code/) are in the repo; the GPU-only steps (training, re-running evals) are documented with exact commands in REPRODUCE.md.

Build the paper

cd paper
pdflatex main.tex && pdflatex main.tex   # bibliography is embedded (thebibliography)

Run the site

cd site
npm install
npm run dev      # local dev
npm run build    # static build → dist/

Citation

@article{li2026userasengram,
  title         = {User as Engram: Internalizing Per-User Memory as Local Parametric Edits},
  author        = {Li, Bojie},
  year          = {2026},
  eprint        = {2606.19172},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2606.19172}
}

License

• Code (this repository — code/, site/, paper/ figure scripts, data/): Apache License 2.0.
• Paper (the manuscript text and figures under paper/): CC BY 4.0, matching the arXiv posting (see paper/LICENSE).

Third-party attribution (nanochat, Engram) is in NOTICE. Copyright © 2026 Pine AI.