Liltloom (语织): A User-Controllable DSH Writing Style Memory Layer Prioritizing Chinese

Liltloom (语织) is a Chinese-first AI writing style memory layer maintained by Adkid-Zephyr under the MIT license, specifically designed for DeepSeek Harness (DSH). It addresses the issues of scattered style requirements, difficulty in review, and lack of portability by transforming style from temporary prompts into structured, editable, and migratable rule data. Core features include silently learning preferences from user expressions to generate bounded Style Context Packets; supporting rule viewing, editing, locking, and JSON export/import; and providing the `/liltloom` command for previewing, enabling, or switching modes. By default, explicit retrieval is employed, with no full conversation texts stored, and Deep Style raw excerpts with small-model analysis are disabled by default. Installation requires Node.js ^22.19.0 or higher, pnpm, and