HiMGA

HiMGA is a Hierarchical Multi-Graph Architecture for conversational memory.

Installation

User

Core install — includes exact_match, F1, ROUGE, BLEU, METEOR:

pip install himga

Full eval metrics — adds BERTScore and Sentence-BERT:

pip install "himga[eval]"

GPU acceleration (BERTScore / SBERT auto-detect CUDA, no code changes needed):

# Install CUDA-compatible PyTorch first, then eval extras
pip install torch --index-url https://download.pytorch.org/whl/cu118
pip install "himga[eval]"

Contributor

Project uses uv for dependency management.

git clone https://github.com/colehank/HiMGA
cd HiMGA
uv sync --dev                  # installs all deps including eval extras
uv run pre-commit install --hook-type pre-commit --hook-type pre-push --hook-type commit-msg

Copy .env_example to .env and fill in the required values.

Running tests:

uv run pytest tests/            # fast tests only (default)
uv run pytest tests/ --run-slow # include BERTScore / SBERT model tests
uv run pytest tests/ --run-integration  # include real API calls (requires .env)

Releasing to PyPI:

git tag vx.x.x
git push origin vx.x.x          # triggers build + publish to PyPI

Benchmarks

HiMGA is evaluated on locomo and longmemeval. Set DATASETS_ROOT in .env (defaults to .cache/datasets) to the directory where datasets exist or should be downloaded.

from himga.utils import get_dataset

locomo   = get_dataset("locomo")       # downloads if not found locally, returns local path
longmem  = get_dataset("longmemeval")