MIT 协议、本地优先的 agent memory 内核，经 Cordis 插件生命周期原生载入 DeepSeek Harness：类型化知识图谱、跨会话自动召回与 gm_* 工具，SQLite 存储并可选向量检索。
DSH 适配
原生运行时
作者声明
安全审计
未审计
最后核验
2026-08-21
许可证
MIT
01它能帮你完成什么？
Give DeepSeek Harness agents durable, traceable cross-session memory
A typed knowledge graph (TASK / SKILL / EVENT nodes with typed edges) in local SQLite, with relevant memory automatically recalled in new sessions — even after DSH restarts
DeepSeek Harness users who want their agent to remember past tasks, fixes and reusable skills across sessions
Shrink context by recalling only relevant memory instead of replaying full history
A relevant local subgraph injected into the prompt during Prompt Assembly; the author's seven-turn benchmark measured an ~75% token reduction at turn R7
DSH users hitting context limits or paying for redundant history replay in long workflows
Explicitly record and search long-term knowledge inside DSH
gm_record persists TASK / SKILL / EVENT nodes deterministically; gm_search runs long-term graph search; gm_status and gm_stats expose store and graph state
DSH users who need deterministic control over what critical knowledge is remembered and why a memory was recalled
02如何接入 DeepSeek Harness？
前置条件
- Node.js `22.19+` or `24+`
- DeepSeek Harness (dsh) with the web profile — the README uses `npx @deepseek-ai/dsh`; local acceptance was tested on DSH `0.1.0-rc.5`
安装步骤
- 01
Clone and build the tarball from source (the beta is not on npm yet): `git clone https://github.com/adoresever/graph-memory.git && cd graph-memory && npm install && npm test && npm run build && npm pack`
$ git clone https://github.com/adoresever/graph-memory.git && cd graph-memory && npm install && npm test && npm run build && npm pack
- 02
Install the generated tarball into the DSH web profile: `npx @deepseek-ai/dsh plugin --profile web add /absolute/path/to/graph-memory-1.6.0-beta.1.tgz`
$ npx @deepseek-ai/dsh plugin --profile web add /absolute/path/to/graph-memory-1.6.0-beta.1.tgz
- 03
Optionally dump the config to confirm registration: `npx @deepseek-ai/dsh --profile web --dump-config`
$ npx @deepseek-ai/dsh --profile web --dump-config
- 04
Start DSH web: `npx @deepseek-ai/dsh web`
$ npx @deepseek-ai/dsh web
- 05
Optional vector retrieval: export `GRAPH_MEMORY_EMBEDDING_API_KEY`, `GRAPH_MEMORY_EMBEDDING_BASE_URL`, `GRAPH_MEMORY_EMBEDDING_MODEL`, `GRAPH_MEMORY_EMBEDDING_DIMENSIONS` before `dsh web`
$ dsh web
验证接入成功
- Confirm that `graph-memory/dsh` is enabled under Settings → Plugins → Plugin list
- Use `gm_status` to check store path, graph counts, vector coverage, mode, and dimensions
03DSH 适配与能力边界
Native DSH plugin loaded by the Cordis plugin lifecycle via the cordis.patch.yml bundle entry — registers gm_* tools, auto-recall during Prompt Assembly, and DSH Credentials access
Typed knowledge-graph memory core
DSH session conversation events→TASK / SKILL / EVENT nodes with typed edges (USED_SKILL, SOLVED_BY, REQUIRES, PATCHES, CONFLICTS_WITH) plus episodic provenance, stored in local SQLite
Writes a local SQLite database at $DSH_HOME/graph-memory/graph-memory.db (normally ~/.dsh/graph-memory/graph-memory.db)Dual-path recall with graph ranking
the current user query during Prompt Assembly→a deduplicated relevant local subgraph (vector or FTS5 search + community expansion + Personalized PageRank) injected into the prompt
Optional network calls to OpenAI-compatible embedding providers (DashScope, OpenAI, local) when vector retrieval is configuredgm_* tools for record, search and observability
explicit agent or user tool calls inside DSH→gm_record persists knowledge deterministically; gm_search runs long-term graph search; gm_status and gm_stats report store, extraction, recall, vector and community state
gm_record writes new nodes into the local graph storeSafe, local-first context assembly
recalled memory nodes→recalled history injected as untrusted reference material that cannot override current user instructions
Memory data stays in the user's local profile by default; API keys come from host credentials or environment variables, never the database
04适合谁？何时不该用？
适合
- DeepSeek Harness users who want their agent to remember past tasks, fixes and reusable skills across sessions
- DSH users hitting context limits or paying for redundant history replay in long workflows
- DSH users who need deterministic control over what critical knowledge is remembered and why a memory was recalled
不适合
- The current build is 1.6.0-beta.1 and DeepSeek Harness is still in Developer Preview, which may introduce compatibility-breaking changes; local acceptance was tested on DSH 0.1.0-rc.5.
- npm registry publication is pending, so installation requires cloning the repository and building the tarball from source instead of a one-command npm install.
- Automatic extraction depends on auxiliary-model output stability, so critical knowledge should be persisted explicitly with gm_record; DSH does not yet expose gm_update and gm_maintain.
05兼容性、维护与安全提示
- The current build is 1.6.0-beta.1 and DeepSeek Harness is still in Developer Preview, which may introduce compatibility-breaking changes; local acceptance was tested on DSH 0.1.0-rc.5.
- npm registry publication is pending, so installation requires cloning the repository and building the tarball from source instead of a one-command npm install.
- Automatic extraction depends on auxiliary-model output stability, so critical knowledge should be persisted explicitly with gm_record; DSH does not yet expose gm_update and gm_maintain.
MIT · actively maintained (last push 2026-08-14; current beta 1.6.0-beta.1)
06常见问题
Graph Memory 如何接入 DeepSeek Harness？原生还是 MCP？
原生接入。它通过 cordis.patch.yml 由 DSH/Cordis 插件生命周期加载，打通 Session、Tool、Agent Loop、Prompt Assembly、LLM 与 Credentials 接缝，不是走 MCP 旁路模拟，也不需要 fork DSH。
安装前需要什么？
Node.js 22.19+ 或 24+，以及 DeepSeek Harness。当前 beta 尚未发布到 npm，需要克隆仓库执行 `npm install`、`npm test`、`npm run build`、`npm pack`，再用 `npx @deepseek-ai/dsh plugin --profile web add` 安装生成的 tarball。
必须配置 embedding API key 吗？
不需要。Embedding 是可选项，不配置时召回自动回退到 FTS5 词法检索。如需向量检索，设置 GRAPH_MEMORY_EMBEDDING_* 环境变量即可接入任意 OpenAI 兼容服务（DashScope、OpenAI 或本地 provider）。
记忆数据存在哪里？
本地 SQLite 数据库：$DSH_HOME/graph-memory/graph-memory.db（通常为 ~/.dsh/graph-memory/graph-memory.db）。数据默认保存在用户本地 profile，API key 来自宿主凭据或环境变量，不会写入数据库。
目前有哪些限制？
当前版本为 1.6.0-beta.1，且 DeepSeek Harness 仍处于 Developer Preview，可能出现破坏兼容性的变更；DSH 侧暂未暴露 gm_update 和 gm_maintain，Pro 可视化图谱工作台也尚未发布。
07相关的 DSH 工作流
weknora
作者 tencent
开源的 LLM 知识库平台：把原始文档转化为可检索的 RAG 知识库、自主推理智能体和自动维护的 Wiki 系统，支持多租户部署。
honcho
作者 plastic-labs
用于构建有状态 AI 智能体的记忆库，支持跨会话连续学习与上下文工程。
mirage
作者 strukto-ai
全球首个面向 AI 智能体的统一虚拟文件系统。
reme
作者 agentscope-ai
ReMe：面向 AI Agent 的记忆管理套件——帮 Agent「记住我、优化我」，支持 RAG 与长期记忆。
08数据与来源
Loaded by the DSH/Cordis plugin lifecycle, not simulated through an MCP side channel.
Integrates Session, Tool, Agent Loop, Prompt Assembly, LLM, and Credentials seams.
页面基于项目公开文档、仓库元数据和 DSH Plugins 的结构化解析生成；最后核验于 2026-08-21。发现错误？提交更正。
最佳 DeepSeek Harness 插件
从全目录挑出的 12 个值得优先安装的插件，覆盖各个分类。