What it is
Reproduction/test repo for SoMem: Can LLM Agents Organize Their Own Memory? (Fu et al., CityU HK + Tencent Jarvis Lab, arXiv:2611.11929). Official code: Qi202/SoMem (MIT).
Independent, unofficial reimplementation. No affiliation with the paper authors. Per our search (2026-11-15, see NOTES.md) this is the first third-party repro of the mechanism.
How it works
The official repo is a full framework (Python 3.12+, PyTorch/sentence-transformers, two paid model endpoints, full-benchmark downloads, and an explicit disclaimer that manuscript scores have not been rerun with the public recipes). This repro implements the core mechanism in a stdlib-only package that is fully offline-testable, and wires a single-endpoint LLM path for real runs on a LongMemEval-S subset. See NOTES.md for the full existing-implementation check.
Try it
git clone https://github.com/Gjusev/somem-repro.git
cd somem-repro
python3 -m unittest # 31 tests, zero dependencies, no network
python3 eval/run_eval.py --fixture data/fixtures/tiny_longmemeval.json --quiet

