← 返回 Skills 市场

Engram MCP Server

Name: Engram MCP Server
Author: dazhuangjammy

作者大壮/Jammy · GitHub ↗ · v0.1.0

cross-platform ⚠ suspicious

447

总下载

当前安装

版本数

在 OpenClaw 中安装

/install engram-mcp-server

功能描述

Engram MCP Server — 给 AI 注入可切换的专家记忆（谁 + 知道什么 + 怎么思考）。支持 Claude Code / Cursor / Codex 等 MCP 客户端。

使用说明 (SKILL.md)

Engram MCP Server

把专家记忆封装成可加载资源：不仅是知识检索，还包含角色、工作流、规则与动态记忆。

适用场景

你希望 AI 在不同专家人设间快速切换
你需要跨会话持久化用户偏好与关键事实
你想在 MCP 客户端里统一管理 memory / knowledge / examples

快速安装（推荐）

claude mcp add --scope user engram-server -- uvx --from git+https://github.com/DazhuangJammy/Engram engram-server

常用命令

# 启动 MCP Server
engram-server serve

# 启动可视化 UI
engram-server ui

关键能力

零向量依赖（不强绑 chromadb/litellm）
索引驱动分层加载（role/workflow/rules 常驻 + knowledge/examples 按需读）
动态记忆、工具轨迹、记忆纠错与归档
支持多种 MCP 客户端集成

参考

GitHub: https://github.com/DazhuangJammy/Engram
文档：README.md

安全使用建议

The skill looks like a legitimate local MCP server for 'engram' memories, but it has several operational risks you should weigh before installing: 1) The installer is a download-and-execute shell command (curl | sh). Inspect https://astral.sh/uv/install.sh yourself before running, or install 'uv' from a package manager you trust. 2) The skill is designed to automatically persist user memory and tool-trace data to project/global folders (./.claude/engram and ~/.engram). That can include PII or sensitive conversation content — avoid loading highly sensitive data into it. 3) The README claims the server will auto-pull the latest code from GitHub on client start; consider disabling auto-update or running the server in an isolated environment if you want reproducible behavior. 4) Review the bundled source code locally (you already have it in the skill bundle) to confirm there are no hard-coded external endpoints or unexpected network calls. 5) If you proceed, run initial installs and the web UI in a sandbox or container, back up any data, and do not enable global automatic installs/updates until you trust the upstream. If you want, I can (a) point out where to inspect the uv installer script, (b) search the included source for outbound-network calls or telemetry functions, or (c) suggest a safer install/run checklist.

功能分析

Type: OpenClaw Skill Name: engram-mcp-server Version: 0.1.0 The skill bundle is classified as suspicious primarily due to its installation method and the ability to install Engram packs from arbitrary Git URLs. The `SKILL.md` file instructs the OpenClaw agent to execute `curl -LsSf https://astral.sh/uv/install.sh | sh`, which is a high-risk practice as it pipes a remote script directly to a shell, creating a supply chain vulnerability. Additionally, the `install_engram` tool allows the agent to `git clone` from user-provided `source` URLs. While the code includes path traversal protections (`loader.py`, `server.py`) to prevent writing outside designated Engram directories, cloning from an untrusted Git repository could still introduce and execute malicious code within the skill's operational scope. There is no evidence of intentional malicious behavior like data exfiltration to unknown endpoints or persistence mechanisms, and the prompt instructions (`CLAUDE.MD`, `README.md`) are aligned with the stated purpose of managing 'expert memory packs'.

能力评估

ℹ Purpose & Capability

The skill's name/description (an MCP server for switchable expert memories) matches the included source files and CLI examples. Requiring the 'uvx' launcher is coherent with the provided install/run patterns (uvx used to run the repo from git). Having full Python source bundled is consistent with the server capability.

⚠ Instruction Scope

Runtime instructions and bundled CLAUDE.md/KNOWLEDGE_VERIFICATION_PROMPT encourage the agent to proactively call MCP tools, automatically capture and persist many forms of user memory (including PII/preferences/tool traces), write MCP config into user/global locations, and prefer running commands itself rather than asking the user. That broad, automatic data-collection/writing behavior expands the skill's scope beyond simple read-only memory lookup.

⚠ Install Mechanism

The SKILL.md metadata includes a shell install step that runs 'curl -LsSf https://astral.sh/uv/install.sh | sh' (download-and-execute). Even if this is the official uv installer, curl|sh is higher-risk than a reviewed package. The skill also documents automatic GitHub pulls on client startup, which means remote code may be fetched and executed without further user action.

ℹ Credentials

The skill does not request environment variables or credentials (no API keys), which is appropriate. However, it is designed to persist user memories (project-level './.claude/engram' and optional '~/.engram') and to record tool traces and memory items automatically; these persistent storage choices can capture sensitive personal data even though no secrets are requested.

⚠ Persistence & Privilege

The install/usage guidance says it will write MCP configuration (potentially global), create project directories, bootstrap starter packs, and auto-pull updates from GitHub on each client start. While 'always' is false, the described global config writes and auto-update behavior grant substantial persistence and the ability to change runtime behavior without repeated user consent.

如何使用

确保已安装 OpenClaw（本地或 Docker 部署）
在对话框中输入安装命令：/install engram-mcp-server
安装完成后，直接呼叫该 Skill 的名称或使用 /engram-mcp-server 触发
根据 Skill 的参数说明提供必要输入，即可获得结构化输出

版本历史

v0.1.0

Initial ClawHub release: Engram MCP Server with expert-memory packaging, index-driven loading, and MCP integration docs.

元数据

Slug engram-mcp-server

版本 0.1.0

许可证 —

累计安装 1

当前安装数 1

历史版本数 1

常见问题