记忆

工程开发 社区

技能解读

当前原文的自动摘译,尚未人工审校。

工作流程

选择集成界面:将@memories.sh/core用于后端路由、worker、cron 作业和非 AI-SDK 代理。将@memories.sh/ai-sdk用于generateText、streamText、中间件或基于ai构建的工具循环。如果任务与 CLI 或 MCP 配置有关,请切换到memories-cli 或memories-mcp。在编写代码之前设置范围:

决策指南

需要从您自己的后端代码直接类型化访问:使用MemoriesClient 需要自动上下文注入到提示或消息中:使用memoriesMiddleware 需要模型通过工具显式读取或写入内存:使用memoriesTools 需要管理存储的技能文件或过程片段:使用client.skills.* 或 AI SDK 技能文件工具 需要租户、密钥或嵌入使用管理:使用client.management.*

参考文件

references/core.md:直接客户端方法、传输选择、错误、管理 API 和技能文件 API references/ai-sdk.md:中间件、工具、预加载、完成后持久性和查询提取模式 references/scoping.md:租户/用户/项目范围规则、服务器端安全性和调试检查表

流狐档案 运行与权限由流狐静态检查推断,不代表作者声明或实际测试结果。
领域
工程开发
作者声明的智能体
未找到明确声明;不据此推断已兼容或已测试
静态检查
94 / 100 · 启发式扫描,不代表运行安全
运行环境证据
未执行;当前没有隔离运行测试证据
作者 / 版本 / 许可证
@webrenew · Apache-2.0
上下文用量评级
低消耗
接入复杂程度
需简单配置
是否需要外部接口密钥
需要 · Vendor-specific
兼容的系统
未声明(默认跨平台)
运行环境要求
未检出明确要求
检测到的文件与系统行为
  • 只读
  • 允许写入 / 修改
网络行为
允许外网请求
安装命令数
无(仅作为资料)

需要注意: 未声明工具权限范围;实际权限由运行环境决定。

作者说明文件(原文) SKILL.md

# memories-sdk

Use the SDK packages when an application needs memories.sh programmatically. Prefer `@memories.sh/core` for direct typed API access and `@memories.sh/ai-sdk` only when the caller already uses the Vercel AI SDK.

## Workflow

1. Pick the integration surface:
   - Use `@memories.sh/core` for backend routes, workers, cron jobs, and non-AI-SDK agents.
   - Use `@memories.sh/ai-sdk` for `generateText`, `streamText`, middleware, or tool loops built on `ai`.
   - If the task is about the CLI or MCP configuration, switch to `memories-cli` or `memories-mcp`.
2. Set scope before writing code:
   - Keep `MEMORIES_API_KEY` server-side.
   - `tenantId` selects the tenant or workspace database.
   - `userId` narrows memory to a user inside that tenant.
   - `projectId` narrows reads and writes to a product area, repo, or feature slice.
3. Use the narrowest pattern that solves the task:
   - Direct CRUD or context lookup: `MemoriesClient`
   - Automatic prompt injection: `memoriesMiddleware`
   - Agent loops with explicit memory tools: `memoriesTools` and `memoriesSystemPrompt`
   - Fetch once and reuse: `preloadContext`
   - Persist after completion: `createMemoriesOnFinish`
4. Verify the integration:
   - Confirm the same scope is used on both reads and writes.
   - Catch `MemoriesClientError`.
   - Do not expose the API key to browser-only code.

## Quick Start

### `@memories.sh/core`

```ts
import { MemoriesClient } from "@memories.sh/core"

const client = new MemoriesClient({
  apiKey: [REDACTED:credential],
  tenantId: "acme-prod",
  userId: "user_123",
})

const context = await client.context.get({
  query: "billing architecture",
  projectId: "dashboard",
  mode: "all",
  strategy: "hybrid",
  limit: 8,
})

await client.memories.add({
  content: "Enterprise billing uses Stripe invoices.",
  type: "fact",
  projectId: "dashboard",
  tags: ["billing"],
})
```

### `@memories.sh/ai-sdk`

```ts
import { generateText, stepCountIs, wrapLanguageModel } from "ai"
import { openai } from "@ai-sdk/openai"
import {
  memoriesMiddleware,
  memoriesSystemPrompt,
  memoriesTools,
} from "@memories.sh/ai-sdk"

const model = wrapLanguageModel({
  model: openai("gpt-4o"),
  middleware: memoriesMiddleware({
    tenantId: "acme-prod",
    userId: "user_123",
    projectId: "dashboard",
  }),
})

const result = await generateText({
  model,
  system: memoriesSystemPrompt(),
  tools: memoriesTools({
    tenantId: "acme-prod",
    userId: "user_123",
    projectId: "dashboard",
  }),
  stopWhen: stepCountIs(5),
  prompt: "Summarize prior decisions about billing.",
})

console.log(result.text)
```

## Decision Guide

- Need direct typed access from your own backend code: use `MemoriesClient`
- Need automatic context injection into prompts or messages: use `memoriesMiddleware`
- Need the model to read or write memory explicitly through tools: use `memoriesTools`
- Need to manage stored skill files or procedure fragments: use `client.skills.*` or the AI SDK skill-file tools
- Need tenant, key, or embedding usage administration: use `client.management.*`
- Need internals of the memories monorepo or server endpoints: use `memories-dev`

## Reference Files

- `references/core.md`: direct client methods, transport choices, errors, management APIs, and skill-file APIs
- `references/ai-sdk.md`: middleware, tools, preload, post-finish persistence, and query extraction patterns
- `references/scoping.md`: tenant/user/project scoping rules, server-side safety, and debugging checklist

讨论

基于 GitHub Discussions。登录 GitHub 即可参与讨论、点赞、订阅更新。