专有 EULA（私有 Beta）。面向 AI 编码 Agent 的结构化 Agent Memory：双时态知识图谱，以 DeepSeek Harness 插件形式经 dsh CLI 安装，本地索引、25+ MCP 工具、17 个技能、零 LLM 调用。
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
NOASSERTION
01它能帮你完成什么？
Give DeepSeek Harness agents persistent structural agent memory of the codebase across sessions
A live bi-temporal knowledge graph of every function, class, call edge and version, queryable in milliseconds via 25+ MCP tools and 17 agent skills
Developers running coding agents on DeepSeek Harness who want shared, replay-aware code context without agents re-reading files every session
Assess blast radius and replay refactors with full causal awareness before changing code
Impact analysis with risk rating (`get_impact`), diff-to-symbol scope mapping (`detect_changes`), and six-mode temporal evolution queries
Engineers doing refactors, incident investigation or code review on large repos who need to know what breaks before they change it
Index a large codebase locally with zero LLM calls and zero API cost
A 50k-file repo indexed in under 90 seconds by Rust + Tree-sitter parsers — 20+ languages plus framework-aware scanners, fully local
Teams with large monorepos or strict privacy requirements who can't send source code through LLM APIs
02如何接入 DeepSeek Harness？
前置条件
- Node.js ≥ 18 (README requirements table)
- DeepSeek Harness CLI — `@deepseek-ai/dsh`, which provides the `dsh` command (installed globally or run via npx)
- Git repository history available — required for temporal analysis
- Private beta access — Memtrace rolls out access in batches via the waitlist at memtrace.io
安装步骤
- 01
Install DeepSeek Harness: `npm install -g @deepseek-ai/dsh`
$ npm install -g @deepseek-ai/dsh
- 02
Add the Memtrace plugin: `dsh plugin --profile web add github:syncable-dev/dsh-plugin-memtrace` (or the npx variant without a global CLI)
$ dsh plugin --profile web add github:syncable-dev/dsh-plugin-memtrace
- 03
Optional: pin a local binary with `npm install -g memtrace` and `MEMTRACE_BIN=memtrace` so the first launch doesn't fetch it via npx
$ npm install -g memtrace
- 04
Ask the agent to index the workspace, then pull blast radius, evolution, or an architecture briefing
验证接入成功
作者未说明
回滚
- `memtrace uninstall` — removes skills, MCP server, plugin, settings
- `npm uninstall -g memtrace`; if npm uninstall already ran, the cleanup script is at `~/.memtrace/uninstall.js`
03DSH 适配与能力边界
Installed as a DeepSeek Harness plugin via the dsh CLI (`dsh plugin --profile web add github:syncable-dev/dsh-plugin-memtrace`); the bundle registers Memtrace's skills and starts `memtrace mcp` inside the Harness profile.
Structural local indexing
any codebase — 20+ programming languages plus YAML / HCL / JSON / TOML / SQL and framework-aware scanners (Express, NestJS, FastAPI, Django, GitHub Actions, Terraform, …)→a live knowledge graph with symbols as nodes (functions, classes, interfaces, types, endpoints) and CALLS / IMPLEMENTS / IMPORTS / EXPORTS / CONTAINS edges, built deterministically with Rust + Tree-sitter — zero LLM calls
writes a local graph index / MemDB on disk; the first index is CPU/RAM intensiveBi-temporal engine
the indexed repo plus its Git history→time-travel queries via six scoring algorithms (impact, novelty, recency, directional, compound, overview) — every symbol carries its full version history
25+ MCP tools + 17 agent skills
natural-language requests from the agent (find / who-calls / what-changed / blast-radius / architecture)→hybrid BM25 + semantic search, relationship analysis, graph algorithms (PageRank, Louvain communities), Cypher queries — skills fire automatically based on what you ask
the DSH plugin bundle starts a local `memtrace mcp` server and registers skill files inside the Harness profileImpact analysis & cross-repo API topology
a symbol, diff, or service boundary→blast radius with risk rating, dead-code detection, complexity hotspots, and the HTTP call graph between repositories
network traffic limited to license validation, aggregate node/edge counts and opt-out crash telemetry — no source code, file paths or symbol names
04适合谁？何时不该用？
适合
- Developers running coding agents on DeepSeek Harness who want shared, replay-aware code context without agents re-reading files every session
- Engineers doing refactors, incident investigation or code review on large repos who need to know what breaks before they change it
- Teams with large monorepos or strict privacy requirements who can't send source code through LLM APIs
不适合
- Memtrace is in private beta — access is rolled out in batches via the waitlist at memtrace.io, so new users may need to wait for a cohort before they can use it.
- Proprietary EULA (GitHub reports NOASSERTION): the indexer and MemDB database are closed-source; free for individual developers during beta and after GA.
- The first index is CPU/RAM intensive — minimum 4 cores, 8 GB RAM, 5 GB disk and Node.js ≥ 18; 8+ cores and 16–32 GB RAM recommended for large monorepos.
05兼容性、维护与安全提示
- Memtrace is in private beta — access is rolled out in batches via the waitlist at memtrace.io, so new users may need to wait for a cohort before they can use it.
- Proprietary EULA (GitHub reports NOASSERTION): the indexer and MemDB database are closed-source; free for individual developers during beta and after GA.
- The first index is CPU/RAM intensive — minimum 4 cores, 8 GB RAM, 5 GB disk and Node.js ≥ 18; 8+ cores and 16–32 GB RAM recommended for large monorepos.
Proprietary EULA per README (GitHub reports NOASSERTION) · actively maintained (latest release v1.1.5, 2026-08-20)
06常见问题
Memtrace 如何接入 DeepSeek Harness？
先安装 Harness CLI（npm install -g @deepseek-ai/dsh），再执行 dsh plugin --profile web add github:syncable-dev/dsh-plugin-memtrace。该插件会注册 Memtrace 的技能，并在 Harness profile 内启动 memtrace mcp。
是原生运行还是 MCP？
以 DSH 插件形式安装：插件包会在 Harness profile 中启动本地 memtrace mcp 服务并注册 17 个 Agent 技能，技能按你的提问自动触发，无需手写提示词。
安装需要什么前置条件？
Node.js ≥ 18 与 DeepSeek Harness CLI；时间线分析还需要 Git 仓库历史。另外 Memtrace 处于私有 Beta，可能需要先在 memtrace.io 排队获取访问权限。
代码会上传到云端吗？
不会。索引完全在本地完成，唯一的网络流量是许可证校验、节点/边数量的聚合统计和可关闭的崩溃遥测，不含源码、文件路径或符号名；用 MEMTRACE_TELEMETRY=off 可关闭遥测。
对机器资源有什么要求？
首次索引较耗 CPU/内存：最低 4 核、8 GB 内存、5 GB 磁盘；大型 monorepo 建议 8 核以上、16–32 GB 内存。后续查询和增量重建索引会轻量很多。
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
Memtrace runs as a [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) plugin.
npx -y @deepseek-ai/dsh plugin --profile web add github:syncable-dev/dsh-plugin-memtrace
That bundle registers Memtrace's skills and starts `memtrace mcp` inside the Harness profile. First launch may fetch the…
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。