Apache-2.0 的 LLM 监督式 AI Agent 持久记忆，通过 dsh-mnemon 插件接入 DeepSeek Harness，构成受监督的三层记忆系统。
DSH 适配
兼容
作者声明
安全审计
未审计
最后核验
2026-08-21
许可证
Apache-2.0
01它能帮你完成什么？
Give DeepSeek Harness agents persistent, cross-session agent memory
Long-term memory spaces in a four-graph knowledge store (temporal, entity, causal, semantic), with intent-aware recall and automatic deduplication
DeepSeek Harness users whose agents lose critical decisions across sessions and need durable knowledge that survives context compaction
Share one pool of agent memory across sessions and frameworks
A single local `~/.mnemon` SQLite database readable and writable by Claude Code, Codex, Cursor, OpenClaw, DeepSeek Harness and other supported runtimes
Teams running multiple agent tools who want cross-session, cross-framework knowledge instead of isolated per-tool memory
02如何接入 DeepSeek Harness？
前置条件
- The `mnemon` binary installed on the host — via Homebrew Cask (`brew install --cask mnemon-dev/tap/mnemon`), `go install github.com/mnemon-dev/mnemon@latest`, or from source (`make install`)
- DeepSeek Harness (dsh) with the Web profile running
安装步骤
- 01
Install mnemon on the host: `brew install --cask mnemon-dev/tap/mnemon` (macOS), or `go install github.com/mnemon-dev/mnemon@latest` (macOS / Linux / Windows)
$ brew install --cask mnemon-dev/tap/mnemon
- 02
Add the DSH plugin: `dsh plugin --profile web add dsh-mnemon`
$ dsh plugin --profile web add dsh-mnemon
- 03
Restart your DSH Web profile: `dsh --profile web`
$ dsh --profile web
- 04
Open DSH's Settings → Plugin Config → Mnemon to pick a storage scope, and use the Memory System tab in a session to create or activate memory spaces
验证接入成功
- `mnemon --version` and `mnemon agency --version` confirm the binary installation
回滚
- `mnemon setup --eject` removes all integrations
03DSH 适配与能力边界
Installed as the dsh-mnemon plugin via `dsh plugin --profile web add` after placing the mnemon binary on the host; layers DSH runtime memory, managed project documents and Mnemon long-term memory spaces
Four-graph knowledge store
facts, decisions and insights the host LLM chooses to remember→temporal, entity, causal and semantic edges — not just vector similarity — with importance decay and garbage collection
memory is persisted to a local SQLite database under `~/.mnemon`LLM-supervised memory operations
the host LLM's judgment calls on what to remember, link and recall→structured JSON results from three primitives (`remember`, `link`, `recall`); no embedded LLM, no API keys
`mnemon setup` deploys skill files, hooks and behavioral guides into each runtime's config directoryIntent-aware recall
natural-language recall queries from the agent→graph traversal plus optional vector search (RRF fusion), enabled by default for all queries; `remember` auto-detects duplicates and conflicts
Multi-framework memory sharing
one `~/.mnemon` store→memory shared across Claude Code, Codex, Cursor, TRAE, Qoder, OpenClaw, DeepSeek Harness (via dsh-mnemon) and more; named stores and `MNEMON_STORE` isolate per project
optional local Ollama can be added for enhanced vector+keyword hybrid search
04适合谁？何时不该用？
适合
- DeepSeek Harness users whose agents lose critical decisions across sessions and need durable knowledge that survives context compaction
- Teams running multiple agent tools who want cross-session, cross-framework knowledge instead of isolated per-tool memory
不适合
- Windows supports the core Memory commands, but Agency remains unavailable on Windows until its local authority boundary has native Windows security.
- The DeepSeek Harness integration depends on the separate dsh-mnemon plugin (omdsh-dev/dsh-mnemon), which must be added via the dsh plugin CLI after mnemon is installed on the host.
05兼容性、维护与安全提示
- Windows supports the core Memory commands, but Agency remains unavailable on Windows until its local authority boundary has native Windows security.
- The DeepSeek Harness integration depends on the separate dsh-mnemon plugin (omdsh-dev/dsh-mnemon), which must be added via the dsh plugin CLI after mnemon is installed on the host.
- Mnemon is a young project (repository created 2026-02) still below v1; durable writes go through supervised sub-agents and the storage scope must be picked manually in DSH's plugin config.
Apache-2.0 · actively maintained (latest release v0.2.4, 2026-08-19)
06常见问题
Mnemon 如何接入 DeepSeek Harness？
通过 dsh-mnemon 插件接入。先在宿主机安装 mnemon，然后执行 `dsh plugin --profile web add dsh-mnemon` 并重启 DSH Web profile；再打开 Settings → Plugin Config → Mnemon 选择存储范围，会话中用 Memory System 标签页创建或激活记忆空间。
Mnemon 是 DSH 原生插件还是 MCP 服务？
都不是。Mnemon 采用 LLM 监督模式：一个独立的本地二进制，只提供 `remember`、`link`、`recall` 三个原语，由宿主 LLM 决定记什么、何时遗忘——不内嵌 LLM、不走 MCP、无需 API key。
需要什么前置条件？
宿主机上安装 `mnemon` 二进制——可用 Homebrew Cask（`brew install --cask mnemon-dev/tap/mnemon`）、`go install github.com/mnemon-dev/mnemon@latest` 或源码编译——以及运行 Web profile 的 DeepSeek Harness（dsh）。
我的记忆数据存在哪里？
全部保存在本地 `~/.mnemon` 下的单一 SQLite 数据库（可用 `MNEMON_DATA_DIR` 修改位置）；可选的本地 Ollama 可增强向量+关键词混合检索，数据不会离开你的机器。
其他 agent 能共享同一份记忆吗？
可以。默认所有会话共用同一个 `default` 存储，Mnemon 还支持 Claude Code、Codex、Cursor、TRAE、Qoder、OpenClaw 等运行时；命名存储和 `MNEMON_STORE` 环境变量可按项目隔离记忆。
07相关的 DSH 工作流
deepseek-reasonix
作者 esengine
专为 DeepSeek 打造的终端 AI 编程智能体，围绕前缀缓存稳定性设计，可常驻运行。
phi
作者 pulseaiclub
来自 pi 的编码智能体，支持无限提供方、子智能体、行内编辑与权限门控。
sivtr
作者 ariestar
A unified agent memory workspace for human and agent | 一个统一的agent记忆工作空间
deepseek-harness-docker
作者 runzhliu
社区 Docker 与 Kubernetes 打包，含加固镜像、Compose、Helm、Web UI 与无头 CLI。
08数据与来源
which layers DSH's runtime memory, managed project documents, and Mnemon's long-term memory spaces into one supervised t…
With `mnemon` installed on the host (see Install), add the plugin and restart your DSH Web profile: `dsh plugin --profil…
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。