Installation

From PyPI

pip install ai-engram

This installs everything: torch, tqdm, and transformers — so HuggingFace LLMs and the GPT-2 Conv1D path work out of the box. The distribution is named ai-engram; the import package is engram:

import engram
from engram import EngramEditor, EditorConfig
print(engram.__version__)

With uv (notebook / Jupyter)

uv is a fast installer. Create an isolated environment and add ai-engram plus a Jupyter kernel — handy for running the example notebooks:

uv venv                         # create .venv (Python >= 3.9)
source .venv/bin/activate
uv pip install ipykernel ai-engram transformers torch

ai-engram already pulls torch and transformers; listing them explicitly keeps the notebook environment fully reproducible. Register the kernel so it shows up in Jupyter:

python -m ipykernel install --user --name ai-engram --display-name "Python (ai-engram)"

Requirements

• Python ≥ 3.9
• PyTorch ≥ 2.0
• transformers (installed automatically)

From source

git clone https://github.com/jeakwon/ai-engram
cd ai-engram
pip install -e ".[dev]"        # editable install + pytest

The package uses a src/ layout (src/engram) with the hatchling build backend.

Verify

python -c "import engram; print(engram.__version__)"
pytest tests/test_extraction.py -q      # fast CPU unit tests

The heavy TOFU integration tests are gated behind environment variables and require a GPU + cached models — see TOFU validation.