DHMS (Digital Hyperthymesia Memory Systems) is a research artifact and evaluation framework for analyzing large language model (LLM) behavior under controlled memory perturbations. Unlike traditional retrieval-augmented generation (RAG) systems that treat memory as static external knowledge, DHMS models memory as a dynamic influence field that actively modulates model behavior under structured perturbation. This release provides a fully reproducible snapshot of DHMS v1.0, including the paper, reference implementation, and evaluation artifacts. The system defines a controlled perturbation framework with three experimental regimes:- DHMS-A: baseline condition- DHMS-B: memory perturbation condition- DHMS-C: stabilization condition It introduces behavioral metrics for LLM analysis, including:- behavioral divergence under perturbation- hallucination sensitivity proxy- stability under memory interference- correlation between controlled and observed outputs This Zenodo release includes:- Full paper (main.pdf)- Reference implementation (code/)- Evaluation outputs (evaluation/)- Reproducibility assets (assets/)- Citation and metadata files This version is intended for reproducible research, academic citation, and evaluation benchmarking of LLM behavior systems. DHMS is not a product system or prompt engineering method; it is a measurement framework for studying behavioral phase transitions in language models under controlled perturbation. Version: v1.0
Zhong Huaxinsheng (Thu,) studied this question.