AMP MCP server
Commit Oct 29, 2026Provides a shared memory hub for AI agents to synchronize user preferences and facts across different tools via MCP and REST.
Also known as RewFa-CEO/AMP
- Stars
- 7
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Oct 29, 2026
- License
- MIT
What is AMP MCP?
AMP provides a shared memory hub for AI agents to synchronize user preferences and facts across different tools via MCP and REST.
Summary generated from the project's README
About AMP MCP server
Synchronizing user preferences and facts across different AI tools is what AMP enables for developers and power users. It acts as a centralized memory hub that allows agents like Cursor, Claude, and ChatGPT to share a single, consistent memory. By connecting via the Model Context Protocol (MCP) or REST API, these tools can instantly access information written by other agents, effectively breaking down data silos between separate chat windows and editors.
Users can implement this by running a local daemon that provides MCP endpoints or a REST API on port 3000. The system supports various storage backends including local JSON files, SQLite, and Redis to manage different scales of data. It also includes a Chrome browser extension for background data collection, such as user highlights and reading preferences. This setup ensures that whether you are using a web-based LLM, a local terminal agent, or a code editor, your context remains unified and accessible.
What you can do with it
- 1Unify personal AI contextA developer can ensure that preferences set in a code editor are immediately known by a web-based chatbot.
- 2Automate context managementAn AI agent can autonomously move unimportant information to long-term storage to stay within its context window limits.
- 3Capture browser activityThe system can automatically turn your web highlights and search history into usable memory fragments.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 15, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablerewfa-ceo/amp is public and reachablegithub.com/rewfa-ceo/amp ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-10-29github.com/rewfa-ceo/amp ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubAMP (Agent Memory Protocol)
One memory for every agent — Cursor, Claude, ChatGPT and your browser, all remembering the same you. 让每一个 AI Agent 共享同一段记忆 —— Cursor、Claude、ChatGPT 与你的浏览器,记住同一个你。
AMP (Agent Memory Protocol) is a plug-and-play memory engine for the LLM era. It features "Pluggable Memory" with true Cross-Agent Shared Inheritance — memory written by one Agent can be instantly used by another. It completely breaks data silos, enabling seamless memory consensus across different AI tools.
AMP (Agent Memory Protocol) — 主打“插件化记忆、即插即用”的大模型时代 AI Agent 记忆管理引擎。实现真正的跨 Agent 共享与继承——一个 Agent 存入的记忆,另一个 Agent 拔插即用,彻底打破信息孤岛,让不同 AI 工具之间形成记忆共识。
🌟 Breakthrough Innovations / 颠覆性创新点
AMP is more than just a key-value store; it's a "Memory OS" for the LLM era: AMP 不仅仅是一个简单的 K-V 数据库,它是大模型时代的“记忆操作系统”,具备以下核心特性:
1. Cross-Application Sharing & Implicit Capture / 跨应用共享与隐式捕获 (The "Memory Hub" Extension)
Unlike traditional memory systems restricted to a single chat window (like ChatGPT or Cursor alone), AMP natively provides a Chrome Browser Extension (extension/).
传统的记忆系统受限于单一对话框(如单独的 ChatGPT 或 Cursor),而 AMP 提供原生的 Chrome 浏览器插件 (extension/) 支持。
- Invisible Data Collection / 无感记忆收集: The extension captures user highlights, reading preferences, and frequent searches in the background, transforming them into dynamically weighted memory fragments. (插件在后台隐式捕获用户高亮内容、阅读偏好及高频搜索,将其转化为动态权重的记忆碎片。)
- Global Interoperability / 全局打通: Whether you are using a web LLM, a local terminal Agent (e.g. OpenClaw), or a code editor (e.g. Claude Code), they can all connect to the same AMP memory hub via the MCP (Model Context Protocol). (无论是网页版大模型、本地终端 Agent,还是代码编辑器,都可以通过 MCP 协议或 REST API 接入同一个 AMP 记忆中枢,实现真正的记忆继承与共享。)
2. Autonomous LLM Paging / OS 级的大模型记忆自治
Inspired by modern OS memory paging, we not only provide APIs but also expose getMemoryTools() directly formatted for OpenAI/Claude Function Calling.
借鉴现代操作系统的“内存分页机制”,我们不仅提供了 API,还直接暴露出适配 OpenAI/Claude 的 getMemoryTools() 函数。
- Dynamic Paging / 动态换页: The LLM can autonomously invoke tools to "page out" unimportant context to AMP's long-term memory and "page in" relevant background knowledge when needed, shattering physical Context Window limits. (大模型可以通过主动调用工具,在上下文溢出时自动将不重要的内容“换出”到长期记忆库,并在需要时“换入”关联的背景知识,彻底打破 Context Window 的物理限制。)
3. Hybrid Graph-Vector Architecture / 向量与图谱的双轨混合检索
Pure vector databases struggle with multi-hop logical reasoning. 传统系统单纯依赖向量(Vector DB),难以处理多跳逻辑推理(Multi-hop Reasoning)。
- From day one, AMP aligns
working(Scratchpad),long_term(Vector retrieval), andgraph(Temporal graph memory) tiers in itsMemoryTier. This paves the way for handling deep entity relationships (e.g., "A knows B, and B likes C"). (AMP 从设计之初就在MemoryTier中并列了working、long_term和graph。这为接入 Neo4j 等图数据库,处理“A认识B,且B喜欢C”这类深层实体关系铺平了道路。)
4. Dynamic Weighting & Forgetting Curve / 艾宾浩斯遗忘曲线与动态权重
Every memory entering AMP carries an importance score and lastAccessedAt timestamp. Frequently accessed memories automatically gain weight, while low-value info decays over time (Memory Consolidation).
每一条进入 AMP 的记忆都自带 importance(重要性得分)和 lastAccessedAt(最后访问时间)。高频访问的记忆权重自动增加,低价值信息随时间自动降级(Memory Consolidation),确保检索效率永远保持在最巅峰。
📦 Installation & Quick Start / 安装与快速体验
Install from npm / 从 npm 安装
npm install agent-memory-protocol
import { AMPCore, MemoryTier } from 'agent-memory-protocol';
const amp = new AMPCore();
await amp.store({
tier: MemoryTier.LONG_TERM,
scope: { userId: 'global-user' },
content: 'User prefers writing backend code in TypeScript.',
metadata: { importance: 0.9, tags: ['tech', 'preference'] }
});
const hits = await amp.retrieve({ query: 'code', limit: 3 });
One-line Local Demo / 一行命令跑通本地体验
Run the full memory write + semantic retrieval flow with zero configuration! The system will automatically create a persistent amp_memory.json file for you.
无需任何数据库配置,克隆后直接跑通记忆写入+智能检索的全流程!系统将自动为你创建持久化文件 amp_memory.json,确保断电不丢失。
git clone https://github.com/RewFa-CEO/AMP.git
cd AMP
npm install
npm run build
npm run demo
In the Demo terminal / 在 Demo 终端中:
- Type anything / 直接打字输入 to permanently save your preferences (e.g., "I prefer dark themes for coding" / 即可永久保存你的偏好,如:“我平时写代码最喜欢用深色主题”).
- Type
? theme/ 输入? 偏好to experience cross-Agent semantic retrieval (即可体验跨 Agent 维度的语义检索引擎).
🏗️ Architecture & Ecosystem Support / 架构与生态支持
To satisfy both hardcore developers and full-scenario apps, AMP provides identical core implementations across multiple languages, and they all share the same underlying storage architecture: 为了满足最极客的开发者和全场景应用,AMP 提供完全一致的多语言核心实现,并 共用同一个底层存储架构:
📦 1. TypeScript / Node.js Core (/src)
Best suited for frontend and browser ecosystems. / 最契合前端及浏览器生态的实现。
import { AMPCore, MemoryTier } from 'agent-memory-protocol';
const amp = new AMPCore();
await amp.store({
tier: MemoryTier.LONG_TERM,
scope: { userId: 'global-user' },
content: 'User prefers writing backend code in TypeScript. / 用户偏好使用 TypeScript 编写高并发服务端代码。',
metadata: { importance: 0.9, tags: ['tech', 'preference'] }
});
🐍 2. Python Core (/python)
Tailored for AI Native developers and data scientists. Perfectly integrated with Pydantic type validation. / 为 AI Native 开发者及数据科学家量身定制。完美集成 Pydantic 类型校验。
from amp import AMPCore, MemoryEvent, MemoryTier, MemoryScope, MemoryMetadata
amp_core = AMPCore()
await amp_core.store(MemoryEvent(
tier=MemoryTier.WORKING,
scope=MemoryScope(user_id="global-user"),
content="Developing a disruptive AI memory protocol. / 正在开发一个颠覆性的 AI 记忆协议。",
metadata=MemoryMetadata(importance=0.8, tags=["project"])
))
# Directly get LLM Tools to mount onto GPT-4 or Claude / 直接获取 LLM Tools,挂载给 GPT-4 或 Claude
tools = amp_core.get_memory_tools()
🔌 3. MCP Server & REST API / MCP 记忆中枢与 REST API
Run the global memory hub daemon: / 启动全局记忆守护进程:
npm run start
This single process mounts both standard Model Context Protocol (MCP) endpoints (for Cursor, Claude Code) and HTTP REST APIs (port 3000, for legacy Agents). They both share the identical underlying persistence layer!
该进程将同时提供 MCP (Model Context Protocol) 协议接口(供 Claude Code / Cursor 直接读取)以及暴露在 3000 端口的 HTTP REST API(供其它 Agent 如 OpenClaw 写入)。它们都在底层读写相同的 amp_memory.json / Redis 数据库!
Connect any MCP client in 30 seconds / 30 秒接入任意 MCP 客户端:
Claude Desktop — claude_desktop_config.json (no clone needed / 无需克隆仓库):
{
"mcpServers": {
"amp": {
"command": "npx",
"args": ["-y", "agent-memory-protocol"]
}
}
}
Cursor / Claude Code — .mcp.json (from a local clone / 本地克隆方式):
{
"mcpServers": {
"amp": {
"command": "node",
"args": ["./dist/mcp/index.js"]
}
}
}
Once connected, memories written by one client are instantly retrievable by every other client — that is the whole point of AMP. / 接入后,任一客户端写入的记忆对所有其他客户端即时可见——这正是 AMP 存在的意义。
🎯 Vision / 发展愿景
The ultimate form of AMP is to become the HTTP of the AI Era. Just as web pages fetch resources via HTTP, all future Agents will shake hands via AMP to synchronize human preferences and facts. Users will no longer need to repeat "who I am and what I need" to every new AI that appears.
AMP 的最终形态是成为 AI 时代的 HTTP。就像网页通过 HTTP 获取资源一样,未来的所有智能体(Agents)都将通过 AMP 握手,同步人类的偏好与事实,让用户不再需要向每一个新出现的 AI 重复解释“我是谁,我需要什么”。
🧠 Memory Architecture / 三层记忆架构
| 记忆层级 (Tier) | 说明 (Description) | 用途 (Use Cases) | 存储介质 (Storage) |
|---|---|---|---|
| Working Memory (工作记忆) | 短期高频访问记忆,容量有限 | 当前会话上下文、临时操作记录 | 内存 + 文件 (RAM + File) |
| Long-term Memory (长期记忆) | 永久存储记忆,容量无限 | 用户偏好、项目文档、历史知识 | 本地 JSON / SQLite / Redis |
| Graph Memory (图记忆) | 实体关系存储(规划中) | 知识图谱、关系网络、业务流程(待 Neo4j 集成,见 Roadmap) | 当前作为枚举占位,与 Long-term 共用存储 |
🚀 Performance Metrics / 性能指标
实际性能高度依赖所选后端与数据规模。下表为参考量级(小规模数据 + 默认 FileStorageProvider,单进程基准):
| 指标 (Metric) | 参考性能 (Reference) |
|---|---|
| 单条记忆写入耗时 (Write Latency, File provider) | < 10ms(小数据量) |
| 单条记忆检索耗时 (Retrieval Latency, File provider) | < 50ms(小数据量) |
| 单条记忆检索耗时 (Retrieval Latency, SQLite / Redis provider) | < 10ms(带索引,万级数据) |
| 推荐最大记忆数量 (Recommended Max Memories) | File ~1万;SQLite ~百万;Redis ~百万+ |
| 并发连接支持 (Concurrent Agents via MCP/REST) | 受 Node 单进程限制,约 100–1000 |
⚠️ FileStorageProvider 在每次 retrieve 时会回写
lastAccessedAt,大数据量下应优先选择 SQLite 或 Redis 后端以获得真正的索引/事务能力。
🛡️ Security & Privacy / 安全与隐私
- 100% 本地化: 所有记忆数据默认本地存储,不上传任何第三方服务器。
- At-rest 加密:
FileStorageProvider支持可选的 AES-256-GCM 静态文件加密(通过encryptionKey配置)。注意这是文件级静态加密,传输层加密请通过 HTTPS/TLS 在上层处理。 - 细粒度权限控制: 通过
MemoryScope(userId / sessionId / agentId)实现记忆隔离,retrieve 阶段强制过滤,不同 scope 互不可见。 - 审计日志: 配置
auditLogPath后,所有 store / retrieve / update / delete 操作都会按行追加 JSON 审计条目,可追溯。 - MCP/REST 认证: 生产环境(
NODE_ENV=production)启动时强制要求AMP_API_TOKEN环境变量;token 比较使用crypto.timingSafeEqual防时序侧信道;CORS 默认仅允许 localhost 与 chrome-extension 来源。
💡 Use Cases / 应用场景
- 个人极客 (Personal Users):统一管理所有 AI 助手的记忆,让 Cursor、Claude 和网页版 ChatGPT 共享同一个“大脑”,避免重复输入上下文。
- 企业团队 (Enterprise Teams):共享项目知识库,所有团队成员的 AI 助手都可以访问统一的业务上下文和编码规范(通过 scope 隔离按团队切分)。
- AI 应用开发者 (AI Developers):通过暴露的 REST API / MCP 快速为自定义 Agent 增加长期记忆能力,无需从零开发复杂的记忆系统。
- 知识管理 (Knowledge Management):构建企业级知识库,支持按用户/会话维度的语义检索(基于子串+分词匹配,真正的向量检索见 Roadmap)。
⚖️ AMP vs Other Memory Systems / 与其它记忆方案的对比
The agent-memory track is hot (mem0 ~68k★, Letta ~25k★, Zep/Graphiti ~30k★ as of Oct 2026). They are all great — AMP attacks the problem from a different angle: a shared memory hub every agent plugs into, not another SDK your app must embed. Agent 记忆赛道正热(截至 2026-10:mem0 约 68k★、Letta 约 25k★、Zep/Graphiti 约 30k★)。它们都很优秀——AMP 从另一个角度切入:做一个所有 Agent 都能插上的共享记忆中枢,而不是又一个需要嵌入应用的 SDK。
| AMP | mem0 | Letta (MemGPT) | Zep / Graphiti | |
|---|---|---|---|---|
| Form / 形态 | Protocol + Hub (MCP & REST) | SDK + Cloud Platform | Stateful Agent Framework | Temporal Knowledge Graph Service |
| Cross-agent by default / 跨 Agent 原生共享 | ✅ Core design / 核心设计 | Via platform / 经由平台 | Within Letta apps / 限 Letta 生态 | Via API |
| Runs fully local / 纯本地可跑 | ✅ File / SQLite / Redis | Cloud or self-host | Self-host | Self-host |
| MCP server built-in / 内置 MCP | ✅ | ✅ | ✅ | Partial |
| Browser-side capture / 浏览器侧捕获 | ✅ Chrome extension | ❌ | ❌ | ❌ |
| Vector retrieval / 向量检索 | 🔜 Roadmap | ✅ | ✅ | ✅ (graph) |
| Languages / 语言实现 | TS + Python | Python + TS | Python | Python |
AMP is honest about what it does not have yet: no embedding-based vector retrieval (currently substring + token matching) and no graph layer. See the Roadmap below — stars should be earned by trust, not by hype. / AMP 对尚未具备的能力保持诚实:暂无 embedding 向量检索(当前为子串+分词匹配)与图记忆层,详见下方路线图——star 应该靠信任赢得,而不是靠营销。
🗺️ Roadmap / 路线图
当前版本已实现的功能 vs. 规划中(避免宣传与实现脱节):
| 能力 | 当前状态 |
|---|---|
| File / SQLite / Redis 三种存储后端 | ✅ 已实现 |
MemoryScope 作用域隔离 |
✅ 已实现 |
| 艾宾浩斯遗忘曲线 / 动态权重(半衰期 + 访问增益) | ✅ 已实现 |
| AES-256-GCM 静态文件加密 | ✅ 已实现 |
| 审计日志 | ✅ 已实现 |
| MCP + REST API 双协议 | ✅ 已实现 |
| Chrome 浏览器插件(高亮 / 阅读偏好 / 频繁搜索捕获) | ✅ 已实现 |
| 真正的向量检索(embedding + cosine similarity) | 🔜 规划中(当前为子串 + 分词匹配,无 embedding) |
| Neo4j 图记忆层 / 多跳推理 | 🔜 规划中(MemoryTier.GRAPH 当前仅作枚举占位) |
| 分布式 / 多进程并发安全 | 🔜 规划中(File 路径单进程;SQLite/Redis 路径具备基础并发能力) |
📄 License / 许可证
MIT License - Open Source for the AI Community. / 开源,为 AI 社区而生,完全免费,可自由修改和商用。
AMP MCP questions
What makes AMP different from other memory systems?
Unlike SDKs that must be embedded in specific apps, AMP acts as a shared memory hub that multiple agents can plug into via MCP or REST.
Does it support different programming languages?
Yes, AMP provides core implementations in both TypeScript/Node.js and Python.