@everme/memory-mcp

中文文档
0.7.10发布于 2026/11/06Apache-2.0ES Module近 30 天下载 1963

EverMe memory plugin — generic MCP server for any MCP host (Cursor, Cline, generic JSON-RPC clients). Hosts with native plugin formats (Claude Code, OpenClaw) have their own packages.

原始使用文档
文档目录6 节
  1. Tools exposed
  2. Wire it into a host's mcpServers config
  3. Hosted HTTP mode
  4. Architecture
  5. Tests
  6. License

@everme/memory-mcp

Generic MCP server for the EverMe memory backend. It supports local stdio hosts, hosted stateless Streamable HTTP, and authenticated HTTP+SSE for compatibility with cloud MCP clients such as Manus. Use this from any MCP-speaking host that doesn't have a native EverMe package — e.g. Cursor, Cline, or a generic JSON-RPC client.

If you're on:

  • Claude Code → use @everme/claude-code (native plugin: hooks + commands + skills + this MCP server)
  • OpenClaw → use @everme/openclaw (ContextEngine plugin)
  • Anything else MCP → use this package

Tools exposed

mem_context        durable user Profile snapshot only (no search, no episodes)
mem_search         hybrid memory search (text + vector), optionally bounded
                   to a conversation time window
mem_skill          full text of one agent skill, by the id printed in its
                   search-result line (served from the local skill cache)

Each tool's MCP description carries its own trigger so the host LLM calls it autonomously (load context at session start, recall on a back-reference, save a stated fact) without the user having to ask. The server also advertises the same protocol via MCP instructions for hosts that surface it. mem_search is guided to keep its query short — a few keywords, not the whole conversation pasted in — and to use the default topK (10), so requests stay small and search quality stays high.

mem_search also takes optional startTime / endTime (ISO 8601), which compile into the gateway's filters.timestamp as gte / lte. Either bound may stand alone. They bound when the conversation happened, not when the subject of it happened, and the description tells the model to omit them unless the user named a window that resolves to exact dates — a wrong bound returns zero rows, which reads exactly like a topic that was never discussed. A bound with no offset is completed in the host machine's own timezone, so the model never has to guess where the user is. The applied window is echoed back in the result so a miscomputed date is visible in the same response.

Wire it into a host's mcpServers config

{
  "mcpServers": {
    "everme-memory": {
      "command": "npx",
      "args": ["-y", "@everme/memory-mcp"],
      "env": {
        "EVERME_API_BASE": "https://api.everme.evermind.ai",
        "EVERME_AGENT_ID": "agt_...",
        "EVERME_AGENT_TOKEN": "evt_..."
      }
    }
  }
}

evercli plugin install <host> writes this for hosts without a native plugin package.

Hosted HTTP mode

The package also exposes everme-memory-mcp-http, a Node HTTP service for cloud agents that cannot spawn a local stdio process. All MCP protocol endpoints require an EverMe agent token:

Authorization: Bearer evt_...

Start an installed package with:

EVERME_API_BASE=https://api.everme.evermind.ai \
HOST=0.0.0.0 PORT=3000 \
everme-memory-mcp-http

Or run the published executable without a global install:

EVERME_API_BASE=https://api.everme.evermind.ai \
npx -y -p @everme/memory-mcp everme-memory-mcp-http

Hosted mode needs Node.js 20 or later. On Node.js 18, every POST /mcp request fails with crypto is not defined: the MCP SDK's Streamable HTTP transport uses the global Web Crypto object, which Node.js 18 does not expose. The stdio server, GET /sse and POST /messages work on Node.js 18 (tested on 18.20.8, 2026-10-30).

Terminate TLS at the production load balancer or reverse proxy. Do not pass an emk_* account key or a cloud-memory provider key to these endpoints; the client credential is an evt_* EverMe agent token. Hosted mode does not require EVERME_AGENT_ID because the EverMe BFF derives the agent identity from that token.

The service exposes:

EndpointTransportPurpose
POST /mcpStreamable HTTPStateless requests for modern MCP clients
GET /sseHTTP+SSEOpens an authenticated compatibility session for Manus
POST /messages?sessionId=...HTTP+SSESends JSON-RPC messages to the authenticated SSE session
GET /healthHTTPUnauthenticated load-balancer health check

Each Streamable HTTP request creates a fresh MCP server, transport, and EverMe client. SSE sessions retain their own MCP server until disconnect and are bound to the SHA-256 digest of the token that opened them; /messages rejects a different token even when the caller knows the session id. No credential-bearing client is shared across requests, sessions, or tenants. The MCP layer continues to call only EverMe's stable /api/v1/mem/* BFF; cloud-memory V1/V2 routing stays behind the BFF and is invisible to clients.

VariableDefaultPurpose
EVERME_API_BASEhttps://api.everme.evermind.aiEverMe BFF origin
HOST0.0.0.0Listener address
PORT3000Listener port
MCP_HTTP_MAX_BODY_BYTES1048576Maximum JSON request body (1 MiB)
MCP_HTTP_RATE_LIMIT60Requests allowed per token digest per window
MCP_HTTP_RATE_WINDOW_MS60000Fixed rate-limit window in milliseconds
MCP_HTTP_TOKEN_VALIDATION_MAX_PENDING100Maximum concurrent validations for distinct new tokens

Before a new digest enters the limiter, hosted mode validates the token through the stable BFF /mem/capabilities route; failed credentials consume no limiter capacity. Successful validation is cached for the current window, and at most 10,000 validated token digests are tracked in memory. Duplicate validations for one token are coalesced, while concurrent validations for distinct new tokens are capped at 100 by default. A horizontally scaled deployment should also enforce shared per-IP and global limits at ingress.

For Manus, generate a remote-agent credential from the EverMe Connect page, open the Manus MCP JSON importer, and paste the generated configuration:

{
  "mcpServers": {
    "everme": {
      "transport": "sse",
      "url": "https://api.everme.evermind.ai/sse",
      "headers": {
        "Authorization": "Bearer evt_..."
      }
    }
  }
}

The token is returned once. Regenerating the Manus configuration rotates the token and invalidates the previous configuration.

Architecture

Thin adapter on top of @everme/agent-sdk. This package owns the MCP framing (stdio, stateless Streamable HTTP, or HTTP+SSE), tool list, and dispatch. Everything else — the EverMe BFF client, realtime agent-memory writes, search/context calls, retry, and redaction — comes from the SDK.

Tests

npm test

License

Apache-2.0

版本与发布记录

@everme/memory-mcp 共发布 28 个版本,首次发布于 2026/05/12;过去 12 个月发布了 28 个版本。

版本发布日期距今
0.7.10latest2026/11/069 天前
0.7.92026/11/0610 天前
0.7.82026/11/0411 天前
0.7.72026/10/3016 天前
0.7.62026/10/1630 天前
0.7.52026/10/121 个月前

关于 @everme/memory-mcp 的常见问题

@everme/memory-mcp 是什么?

EverMe memory plugin — generic MCP server for any MCP host (Cursor, Cline, generic JSON-RPC clients). Hosts with native plugin formats (Claude Code, OpenClaw) have their own packages.

如何安装 @everme/memory-mcp?

在项目目录运行 npm install @everme/memory-mcp,或使用 pnpm add / yarn add / bun add。它还提供命令行工具 memory-mcp、everme-memory-mcp、everme-memory-mcp-http,可以用 npx memory-mcp 直接运行。

@everme/memory-mcp 支持 TypeScript 吗?

该包的 package.json 没有声明类型入口。需要类型提示时,可以在 npm 上查找社区维护的 @types/everme__memory-mcp 是否存在。

@everme/memory-mcp 支持 ESM 还是 CommonJS?

以 ES Module 形式发布,推荐使用 import 引入。在 CommonJS 项目中,Node.js 20.19+ / 22.12+ 可以直接 require() 不含顶层 await 的 ES Module,更早的版本需要改用动态 import()。

@everme/memory-mcp 需要什么版本的 Node.js?

package.json 的 engines 字段要求 Node.js >=18.0.0。版本不符时 npm 默认只给出警告,开启 engine-strict 后会拒绝安装。

@everme/memory-mcp 可以商用吗?

@everme/memory-mcp 使用 Apache-2.0 许可。宽松许可:允许商业使用、修改和再分发,并附带明确的专利授权;分发时需保留许可与 NOTICE 文件,修改过的文件要注明改动。以上为概括,不构成法律意见,请以许可原文为准。

@everme/memory-mcp 的下载量有多少?

根据 npm 官方下载统计,@everme/memory-mcp 在 2026/10/14 – 2026/11/12 这 30 天内被下载 1,963 次,日均约 65 次。下载量会把 CI 构建、镜像同步和作为其他包依赖的安装一并计入,适合横向比较热度,不能直接等同于使用人数。

@everme/memory-mcp 还在维护吗?

最新版本 0.7.10 发布于 2026/11/06(9 天前);至今共发布 28 个版本,其中过去 12 个月发布了 28 个;首次发布于 2026/05/12。发布频率只是参考信号:功能稳定的工具可能很少需要更新,建议结合源码仓库的 issue 和提交记录判断。