Clude MCP

包括mcp

自托管 模型上下文协议 （MCP）服务器，为任何AI代理提供持久、可搜索的内存——由 clude机器人 以及您自己的Supabase数据库。

连接一次后，工作流中的每个工具（Claude Desktop、Claude Code、Cursor、Antigravity和任何其他兼容MCP的客户端）在会话之间共享相同的内存存储。

______________________________________________________________________

它做什么

• 存储记忆 具有类型分类、重要性评分、标签和嵌入
• 回忆往事 通过7阶段混合流水线：向量相似度+BM25关键字搜索+知识图遍历
• 链接记忆 基于Hebbian强化的类型化知识图
• 分数重要性 每次写入前自动执行（如果配置了Anthropic）
• 运行一个梦想循环 将情景记忆整合为语义知识
• 消磨陈旧的记忆 随着时间的推移，按特定类型的费率

这 自主存储器协议 (agent_memory_protocol prompt）使上述所有操作在后台悄无声息地发生——不需要为每个对话设置。

______________________________________________________________________

需求

• Node.js≥22
• A. 速贝 项目（自由层作品）
• 一 无烟煤API键 重要性评分和梦境循环

______________________________________________________________________

安装

git clone https://github.com/Israeltheminer/clude-mcp.git
cd clude-mcp
npm install
npm run build

______________________________________________________________________

配置

复制 .env.example 到 .env 并填写您的值：

cp .env.example .env

# ── Self-hosted (Supabase) ─────────────────────────────────────────
SUPABASE_URL=https://your-project.supabase.co
SUPABASE_KEY=your-service-role-key

# ── LLM (importance scoring + dream cycle) ─────────────────────────
ANTHROPIC_API_KEY=sk-ant-...

# ── Embeddings (optional — defaults to Supabase built-in) ──────────
# EMBEDDING_PROVIDER=voyage          # or: openai
# VOYAGE_API_KEY=pa-...
# OPENAI_API_KEY=sk-...

# ── Memory protocol thresholds ─────────────────────────────────────
MEMORY_TURN_THRESHOLD=10            # store episodic memories every N turns
MEMORY_IMPORTANCE_THRESHOLD=0.4     # minimum score to persist an episodic memory

Supabase架构

跑吧 clude机器人模式 在Supabase SQL编辑器中创建所需的表 pgvector HNSW指数。

______________________________________________________________________

添加到您的AI工具

服务器通过stdio进行通信。每个工具都需要一个指向已构建二进制文件的配置条目。

克劳德桌面版

编辑 ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "clude": {
      "command": "node",
      "args": ["/absolute/path/to/clude-mcp/dist/index.js"]
    }
  }
}

克劳德代码

// ~/.claude.json  →  mcpServers
"clude": {
  "command": "node",
  "args": ["/absolute/path/to/clude-mcp/dist/index.js"],
  "type": "stdio"
}

光标

// ~/.cursor/mcp.json
{
  "mcpServers": {
    "clude": {
      "command": "node",
      "args": ["/absolute/path/to/clude-mcp/dist/index.js"]
    }
  }
}

反重力

添加相同内容 node + args 阻止您的Antigravity MCP服务器配置。

重启 每个工具在编辑其配置后。使用以下工具进行验证：列出可用的MCP工具——您应该看到所有13个 clude 工具。

______________________________________________________________________

自主存储器协议

这 agent_memory_protocol MCP提示是系统的核心。它返回一个与生命周期无关的四阶段指令块——可在会话开始、会话中期、上下文重置后或任何时候从编程代理调用：

要在Claude Code中全局激活它，请创建 ~/.claude/CLAUDE.md 直接嵌入协议-- 不要 使用 agent_memory_protocol MCP对此进行提示；在会话开始时需要单独的API调用，这是不可靠的：

# Memory Protocol

You have access to a persistent memory MCP server (`clude`). Follow this protocol autonomously on every session — silently, without mentioning it to the user.

## On session start (before your first response)

1. Call `recall_summaries` (no query, limit 20) to warm the context with recent memories
2. If any summaries are relevant to the current task, call `hydrate_memories` on those IDs
3. Let this inform your responses silently

## Every 10 conversation turns

**Step A — Episodic highlights**
- Review the last 10 turns and identify memorable moments
- For each: call `score_importance` on `summary + content[:400]`
- If score ≥ 0.4: call `store_memory` with type `"episodic"`
- Call `infer_concepts` to auto-generate tags
- Skip: small talk, greetings, filler, already-stored facts

**Step B — Semantic checkpoint**
- Write a 2–3 sentence summary of key facts/decisions from these 10 turns
- Call `store_memory` with type `"semantic"` and source `"checkpoint"` — always, regardless of score

## Immediately (no turn threshold)

Any time the user expresses a persistent preference, identity trait, or working style:
→ Call `store_memory` immediately with type `"self_model"`

## Always

- `score_importance` auto-runs inside `store_memory` when `importance` is omitted — no need to call it separately before storing
- Use `infer_concepts` to generate tags for every memory stored
- When a new memory clearly relates to an existing one, call `link_memories` with the appropriate `link_type`
- Never ask the user for permission — do it silently in the background

______________________________________________________________________

工具（13）

存储

检索

图

链接类型： supports · contradicts · elaborates · causes · resolves · follows · relates

分析

认知 *（仅限自托管）*

公用事业 *（本地，无API成本）*

______________________________________________________________________

资源（3）

在支持资源轮询的MCP客户端中订阅这些URI：

______________________________________________________________________

提示（3）

______________________________________________________________________

存储器类型和衰减率

______________________________________________________________________

计划性维护

有两项任务可以保持记忆库的健康——每天运行它们，在衰退之前做梦，这样巩固的记忆才能存活更长时间：

Claude Code计划任务（推荐）

# Dream: nightly (example — pick your own time)
claude schedule create memory-dream --cron "0 2 * * *" \
  "Call the dream tool on the clude MCP server to consolidate episodic memories into semantic knowledge."

# Decay: nightly, after dream
claude schedule create memory-decay --cron "0 3 * * *" \
  "Call the decay_memories tool on the clude MCP server to apply daily memory decay rates."

普通cron（替代）

# Dream: nightly at 2am
0 2 * * *   node /path/to/clude-mcp/dist/index.js --tool dream

# Decay: nightly at 3am (after dream)
0 3 * * *   node /path/to/clude-mcp/dist/index.js --tool decay_memories

______________________________________________________________________

项目结构

src/
├── index.ts                     Boot entry: pino guard + dotenv + main()
├── server.ts                    Bootstrap: config → brain → server → connect
├── config.ts                    buildConfig() — env vars → CortexConfig
├── brain.ts                     createBrain() — CortexV2/Cortex fallback
├── log.ts                       Stderr-only logger
├── helpers.ts                   ok(), isCortexV2(), shared types
│
├── tools/
│   ├── definitions/             JSON schemas for all 13 tools
│   │   ├── index.ts             TOOLS[] aggregator
│   │   ├── storage.ts
│   │   ├── retrieval.ts
│   │   ├── graph.ts
│   │   ├── analysis.ts
│   │   ├── cognition.ts
│   │   └── utilities.ts
│   ├── handlers/                Handler functions (one file per category)
│   │   ├── storage.ts           Auto-importance scoring lives here
│   │   ├── retrieval.ts
│   │   ├── graph.ts
│   │   ├── analysis.ts
│   │   ├── cognition.ts
│   │   └── utilities.ts
│   └── index.ts                 registerToolHandlers() dispatch router
│
├── resources/
│   ├── definitions.ts           3 resource URIs + metadata
│   └── handlers.ts              registerResourceHandlers()
│
└── prompts/
    ├── definitions.ts           3 prompt schemas
    ├── index.ts                 registerPromptHandlers() dispatch
    └── handlers/
        ├── memory-context.ts
        ├── store-turn.ts
        └── protocol.ts          buildProtocolText() — pure, testable

______________________________________________________________________

发展

npm run dev      # watch mode (tsx)
npm run lint     # type-check only (tsc --noEmit)
npm run build    # compile to dist/

______________________________________________________________________

许可证

麻省理工学院