Mnemosyne Memory：零依赖、本地优先的 Agent 记忆 Skill
Mnemosyne Memory（@user_663a53c6/mnemosyne-memory）是 SkillHub 上的 ai-agent 类 Skill，当前页面版本 v7.0.1。官方定位为 Mnemosyne OS：仅依赖 Python 3.8+ 标准库、本地优先，无需向量库与独立 LLM 运行时，支持多层次遗忘、哈希链审计、记忆图谱、插件 SDK、本地 Web 管理与 MCP。可作为 Python 库、CLI、HTTP API、MCP server 或 stdio 嵌入 Agent。本文基于 SkillHub 公开资料说明安装命令（skillhub install 与 Cursor/Claude/Codex skills 目录）、接入形态与适用注意，并给出官方页面链接。
Read Moredsh-memory-local：DSH 本地跨会话记忆插件
dsh-memory-local 是 LISHUANGXI1412 维护的 DSH 社区插件（MIT，0.1.1），在本机 ~/.dsh/memory/ 存储跨会话记忆，提供 memory_remember 等五项 Agent 工具与有界 system prompt 注入，支持 project/global 作用域与 Settings→Memory 管理页。安装方式为 plugin_manager install_bundle 指向本地检出目录。无网络、无 embedding，显式写入并过滤疑似凭证内容。
Read Moredsh-evolution：本地会话账本与跨会话复盘
dsh-evolution 是 wycto 维护的 DSH 记忆类插件：将会话结构化写入本地 NDJSON 账本，可选地用已配置的模型做跨会话复盘，将带证据链的洞察注入同仓库后续会话。默认不联网、不上传、不写代码仓库；自动复盘与外向能力均默认关闭。本文基于官方 README 介绍安装命令、数据目录、面板用法与隐私默认值。
Read MoreAI Agent记忆系统：本地优先的跨会话长期记忆 Skill
AI Agent记忆系统（@user_7db0c006/agent-local-memory-keeper，v5.2.1）是 SkillHub 上的 ai-agent 类 Skill，面向跨会话失忆与结论过时问题。对话中的踩坑、纠偏与验证结论可自动沉淀；库内去重、纠错、冲突仲裁与定期自检；支持自然语言语义召回（本地中文模型、离线、数据不出本机）、版本回滚与时间点回溯。安装需 skillhub CLI 并用 --dir 指向 Cursor 等 Agent 的 skills 目录。
Read MoreMyLife: A Local-First Personal State Profile and Decision Engine
MyLife is a local-first memory plugin for DeepSeek Harness (DSH), designed to address challenges such as maintaining agent identity and state, extracting key threads from messy information, and self-questioning during decision making. The plugin includes a four-layer information model, timeliness checks, a counter-evidence engine, and provenance tracking, supporting the invalidation of outdated premises. Its core features are local-first operation, data not leaving the local store, and non-reading of credentials, under the MIT license. After installation, users can interact directly in natural language; data is stored locally only and does not involve network requests.
Read Moredsh-web-search-ddg: Keyless DuckDuckGo Search for DeepSeek Harness
`dsh-web-search-ddg` is a web tool plugin for DeepSeek Harness that provides agents with free local search capabilities by scraping DuckDuckGo Lite pages. The plugin features zero credentials, zero token consumption, strict parsing, and automatic deduplication, and supports GET and POST requests. After installation, it can be used by specifying the search provider in the configuration, making it suitable for developers who prioritize privacy and require no API keys or audit logic.
Read Moreming-xia/dsh-stickies: A Local-First Markdown Sticky Notes Plugin for DeepSeek Harness
`ming-xia/dsh-stickies` is a local Markdown sticky notes plugin designed for DeepSeek Harness. It integrates sticky notes into the conversation page, supporting draggable floating cards, Markdown editing (including task lists, code blocks, and more), and color management. All data is stored locally, with no cloud account required. Users can install it via CLI and use `node scripts/serve-local.mjs` for local preview. This plugin helps developers quickly capture context and drafts without switching windows.
Read Moredsh-think-translate: A Display Layer Translation Plugin for the DeepSeek Harness Web UI
dsh-think-translate is a local-first translation plugin for the DeepSeek Harness (DSH) web UI, designed to remove language barriers in model thinking chains, task cards, and responses. It translates source text into eight languages in real time, uses a local Ollama model by default (such as qwen2.5:7b), and supports privacy protection, zero context cost, and code-safe skipping. As a display-layer tool, it only translates frontend content without interfering with model execution, and can automatically fail over to Google or Bing.
Read MoreDeepDeck Video Sherlock: A Local-First Video Forensics Tool
DeepDeck Video Sherlock is a localized video investigation plugin designed specifically for DeepDeck Harness (DSH). By injecting Agent Skills, it provides video retrieval, transcription, semantic indexing, and visual analysis within the App Workspace. The plugin supports local processing of sensitive data and offers native playback, interactive timeline navigation, multidimensional evidence review, and Markdown audit reports. Users can install it through the DeepDeck interface or local development mode, meeting the needs of a closed-loop workflow for secure, deep semantic analysis of video data.
Read Moredsh-opendesigner: A Local-First React/Tailwind Visual Designer
The `dsh-opendesigner` plugin for DeepSeek Harness (DSH) provides developers with a local-first React and Tailwind visual editor, addressing the lack of design tooling in the DSH environment. It enables Agents or developers to manipulate design mockups directly on a constrained tool surface, supporting local editing, approval for destructive writes, and Mock mode. Installation requires Node 22.14+ and a Host restart, while overlays can be used to debug or preview designs. The plugin ensures security through path jails and approval channels, and prohibits API keys from being uploaded to the repository.
Read Moredsh-writing-desk: A Local-First Writing Workbench for DeepSeek Harness
dsh-writing-desk is a local-first writing plugin for DeepSeek Harness, designed to address the context confusion caused by mixing reference materials with main text in long-form writing. It uses a half-window panel to run the left-side chat and the right-side local writing workbench in parallel, providing volume/chapter file management, Markdown editing, source referencing, and multi-format export. The plugin uses atomic writes to ensure data safety, enforces strict limits on file system and network permissions, and is suitable for local writing workflows that combine AI-assisted creation.
Read Morewinyh/dsh-growth: DeepSeek Harness Local-First Growth Analysis and SOP Generation Plugin
`winyh/dsh-growth` is a local-first plugin for DeepSeek Harness designed to address the challenges of transforming fragmented data in formats such as Markdown, CSV, and JSONL. It provides multi-format parsing, growth artifact validation, and attribution analysis capabilities, covering scenarios such as AARRR, retention, and MRR, and can generate actionable diagnostics and SOPs. Users only need to install it into the specified profile and restart the configuration to use it, making it suitable for product managers and growth hackers who want to turn data into executable growth solutions.
Read MoreOpenBook RSS Reader: A Local-First RSS Reading and Knowledge Collection Plugin
OpenBook, a DeepSeek Harness plugin, is a local-first RSS reader refactored as a DSH plugin, offering a three-pane reading interface, note management, and full-text reading capabilities. It runs using native Cordis services, supports environment awareness, can inject article content into the Agent context, and enables discussion, summarization, and note organization for articles via commands and tools. Data is stored in SQLite, and SSRF protection is enabled by default. Once installed via the CLI, it integrates seamlessly into the Web UI, deeply combining knowledge collection with agent workflows.
Read MoreDSH Knowledge Forge: A Local-First Automatic Wiki and Skill Memory Plugin
DeepSeek Harness (DSH) Knowledge Forge is a native plugin designed to provide local knowledge distillation capabilities for DSH Agent sessions. It converts conversation trajectories into progressively reusable knowledge and separates operational evidence from knowledge. Its core features include local-first capture, session-aware batching, one-time distillation, Wiki-to-Skill evolution, progressive disclosure, and white-box observability. It uses SQLite to maintain runtime state and Markdown files to store Wiki pages and native DSH Skills, enabling structured knowledge to be accumulated from conversations and improving reuse efficiency. The plugin requires DeepSeek Harness 0.1.0-rc.6 and Node.js.
Read Moredsh-voice: Voice input to user message, readable replies.
The DSH plugin `dsh-voice` aims to enhance terminal agent efficiency through voice interaction, implementing speech-to-text, text-to-speech, and reading aloud. Adhering to a "local-first" philosophy, the plugin stores audio files locally by default and does not automatically record or upload to the cloud unless users explicitly configure cloud backends such as OpenAI or Edge-TTS. Its core functions include: `transcribe`, which supports transcription from files or real-time microphone recording, inserting results into the session as voice cards; `speak`, which synthesizes and plays text in the background, suitable for announcing long task status; and the `/voice` command, used to toggle session-level auto-read replies (defaulting to off). The plugin is compatible with Node.js >=20, supports local backends like Whisper-local and macOS Say, and can work with `dsh-cr`.
Read MoreDSH-Failure-Capsule: A Local Failure Evidence Bundle Plugin for DSH
`dsh-failure-capsule` is a local failure evidence-package plugin for DeepSeek Harness, maintained by YiHarvest under the MIT license. It addresses the challenge of diagnosing complex failures in DSH scenarios by automatically collecting relevant signals and generating a sanitized ZIP evidence package when errors occur during tool invocation, turn completion, or Agent runtime. The plugin adheres to a local-first principle, with no data upload or remote diagnostics involved. Its core features include: capturing various failure events (e.g., `tool/result` errors, `turn/end` interruptions); default sanitization of credentials, keys, private keys, and sensitive paths; and collecting Git status, runtime information, plugin manifests, and a bounded Session Event timeline (defaulting to a maximum of 80 entries).
Read Morehanai-investment-dsh: A Local-First A-Share Research Workbench Powered by DeepSeek Harness
`hanai-investment-dsh` is a local-first A-share research workbench plugin built on DeepSeek Harness (DSH). It integrates market panorama, custom stock groups, individual stock K-line analysis, valuation display, expert dialogues, and research report generation, aiming to provide a complete research loop from observation, discussion, to verification. Core features include monitoring six major indices and sector heatmaps, customizable stock watchlists, forward-adjusted K-lines with turning point markers, five-dimensional valuation radar, open dialogues among five experts (including an AI-simulated perspective), and Agent-based retrieval verification to generate archivable Markdown research reports. This tool emphasizes evidence chains and context, clearly distinguishing between price-volume hints and trading signals. All backtest data serves solely as historical retrospective evidence, with no promises of returns. It is designed for developers and investors conducting systematic A-share research within the DSH environment.
Read Moredsh-explain: Transforming DSH Work Sessions into a Local-First Learning Cycle
dsh-explain is a DSH learning mode plugin maintained by yuezengwu, compatible with DSH 0.1.0-rc.8. It extracts concepts from completed work sessions, generates structured learning cards, writes to cross-session global learning threads, and tracks learning progress via ExplainContext. The main agent remains unaffected, with data prioritized for local storage in SQLite. It supports manual /explain, selected text explanation, learning from responses, and automatic evaluation, while providing a diagnostic settings interface and context compression.
Read MoreJingling (dsh-jingling): A Local-First DSH Companion Plugin
Jingling is a DSH companion plugin maintained by Yi-111-a, designed for DSH 0.1.0-rc.6, with the current version dsh-jingling@0.1.0-alpha.19. She offers reviewable memories, Character Card V2 import, a web companion panel, and an optional Windows transparent desktop pet sidecar; she does not possess Shell, file system, network, or MCP permissions, and actual operations require switching back to the standard session. Installation is supported via one-click install-all.ps1, npm installation, or manual source packaging.
Read MoreIntegrating a Local-First Long-Form Novel Workbench into DeepSeek Harness with AI-Novel-Writer
AI-Novel-Writer, maintained by EthanYoQ, is a local-first AI novel writing workspace. On the desktop side, long-form assets are organized into five categories: premise, characters, blueprints, drafts and reviews. There is also a development preview of `@ethanyoq/dsh-ai-novel-writer 0.1.0` in the repository, which connects the five types of assets to the DeepSeek Harness Web novel writing workspace, with writing processes going through single-file native approval. The GitHub directory page for the project is github:EthanYoQ/AI-Novel-Writer, where the installation instructions conflict with the repository description: the root package is a GPL-3.0 licensed desktop application, while the activatable DSH bundle is located in the plugin subdirectory (licensed under MIT) and does not read desktop `.vela` projects.
Read More