学而不厌,诲人不倦。——《论语》

Supermemory:给你的 AI 装上一颗“会记事”的心脏(而且跑得飞快)

Supermemory 走进来时,身上写着一句话——像自我介绍,也像宣言:

Memory engine and app that is extremely fast, scalable. The Memory API for the AI era.

它不爱绕弯子:
它就是来解决那件你早就烦透了的事——你的 AI 每次聊完就忘

你刚把偏好、项目背景、关键约束说得明明白白,下一次对话它又像刚开机一样:“你是谁?你想要什么?”
Supermemory 把这件事当成了它的天职:把“每次对话都像第一次见面”的尴尬,变成“我一直都懂你”的默契。


它到底是什么:AI 时代的“记忆与上下文层”

Supermemory 不只是一个“存点笔记”的小工具。它更像一个上下文层(memory + context layer),把 AI 需要的“人设、长期事实、最近状态、知识库检索、外部数据源同步、文件解析”全塞进一个系统里。

它自己说得很清楚:

  • 你 AI 在会话之间会忘东西
  • Supermemory 自动从对话里学、抽取事实、维护用户画像
  • 处理知识更新与矛盾
  • 该忘的也会忘(过期信息自动失效)
  • 并在“正确的时机”把“正确的上下文”递给 AI

一句话:Full RAG + connectors + file processing —— 整套 context stack,一个系统。


它像一个什么性格的搭档?

它不是那种“你问我答”的被动型助手。
它更像一个安静但极其勤快的秘书:你在前台聊天,它在后台默默做三件事——

  1. 记住值得记住的
  2. 忘掉应该忘掉的
  3. 在你需要的时候把“你真正的你”递出来

它很懂分寸:不把噪音当记忆,不把临时当永久,不把新事实和旧事实硬挤在一起。


你能从它身上得到什么:五件大事

Supermemory 在 README 里把自己的能力摆成了表格,我把它“拟人化”翻译一下:

1)Memory:它会从对话里“提炼事实”

它不只保存全文,它会提炼“事实”,并且能处理:

  • 时间变化(你搬家了、换工作了、项目变了)
  • 矛盾更新(新信息覆盖旧信息)
  • 自动遗忘(临时事项过期就淡出)

2)User Profiles:它会维护你的“用户画像”,而且很快

它把“稳定事实 + 最近动态”分层维护。
然后你只要一通电话(一次 API call),它就能把画像塞回给 AI。

它甚至强调了速度:One call, ~50ms。

3)Hybrid Search:它把 RAG 和 Memory 合并成“一次查询”

很多系统是:你要么检索知识库,要么检索记忆。
Supermemory 是:你问一次,它把**知识库文档(RAG) + 个人化记忆(Memory)**一起给你。

4)Connectors:它不满足于你说的话,它还会同步你的世界

它支持把外部数据源自动同步进来(并且有实时 webhook):

  • Google Drive
  • Gmail
  • Notion
  • OneDrive
  • GitHub
  • ……

5)Multi-modal Extractors:它不挑食,PDF、图片、视频、代码都能吃

它能处理:

  • PDFs
  • images(OCR)
  • videos(transcription)
  • code(AST-aware chunking)

你上传,它就开始干活。


两条路:给“自己用的 AI”装记忆,或给“产品里的 AI”装记忆

Supermemory 的 README 直接把使用方式分成两类:

A. “我用 AI 工具”:我想让我的 Claude / Cursor / VSCode 记住我

你可以用 Supermemory 的 app、浏览器扩展、插件、MCP server。

它的目标是:一键安装,你的 AI 就开始记住你。

B. “我在做 AI 产品”:我想给我的 Agent / App 加上记忆、RAG、画像、连接器

你可以直接用它的 API(并且有 npm/pypi 包)。

它的目标是:一个 API,整个 context stack 都给你。


给你的 AI 记忆:MCP 快速安装(命令行一把梭)

Supermemory 很懂现在的生态:你不一定想写代码,你可能只想“装上就能用”。
它给了 MCP 的一条命令:

1
npx -y install-mcp@latest https://mcp.supermemory.ai/mcp --client claude --oauth=yes

它还会提醒你:把 claude 换成你的客户端,比如:

  • cursor
  • windsurf
  • vscode
  • ……

装好之后,你的 AI 会得到三样“可调用的工具”(它们像三个性格不同的小伙伴):

工具 它做什么
memory 存 / 忘 信息;你说出“值得记”的内容时,AI 会自动调用它
recall 按 query 搜记忆,返回相关记忆 + 你的画像摘要
context 在对话开始时把你的完整画像(偏好、最近状态)注入进去;在 Cursor / Claude Code 里直接输入 /context

MCP 手动配置(如果你喜欢亲手拧螺丝)

最简配置:

1
2
3
4
5
6
7
{
"mcpServers": {
"supermemory": {
"url": "https://mcp.supermemory.ai/mcp"
}
}
}

如果你更想用 API key(不用 OAuth):

1
2
3
4
5
6
7
8
9
10
{
"mcpServers": {
"supermemory": {
"url": "https://mcp.supermemory.ai/mcp",
"headers": {
"Authorization": "Bearer sm_your_api_key_here"
}
}
}
}

它怎么工作:像一只在后台值班的“记忆管家”

安装完成后,Supermemory 会在后台持续做事。它的工作流很朴素,但很有效:

  1. 你照常和 AI 聊天:说偏好、说项目、说问题
  2. Supermemory 抽取并存储重要信息:事实、偏好、项目上下文(不是噪音)
  3. 下一次对话:AI 已经“认识你”,知道你在做什么、喜欢什么、之前聊过什么

它还专门强调了“项目隔离”的概念:记忆带 container tag,你可以按项目、按客户、按仓库、按任何你想要的维度把记忆分隔开。


支持哪些客户端?它很社交,朋友很多

README 里列了它支持的客户端:

Claude Desktop · Cursor · Windsurf · VS Code · Claude Code · OpenCode · OpenClaw

并且 MCP server 也是开源的(有 source 链接)。


如果你要做产品:Supermemory API(npm / pip)就位

它对开发者很友好:安装命令简单粗暴。

1
npm install supermemory    # or: pip install supermemory

然后它会用一段 TypeScript quickstart 把你领进门:

TypeScript:把一段对话变成记忆,再一把拿回画像 + 相关记忆

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
import Supermemory from "supermemory";

const client = new Supermemory();

// Store a conversation
await client.add({
content: "User loves TypeScript and prefers functional patterns",
containerTag: "user_123",
});

// Get user profile + relevant memories in one call
const { profile, searchResults } = await client.profile({
containerTag: "user_123",
q: "What programming style does the user prefer?",
});

// profile.static -> ["Loves TypeScript", "Prefers functional patterns"]
// profile.dynamic -> ["Working on API integration"]
// searchResults -> Relevant memories ranked by similarity

它话里话外都在强调一件事:
你不需要自己折腾 embedding pipeline,不需要自己选向量库,不需要自己发明 chunking 策略。

它像在拍着胸口说:
“这些我都给你包了,你只管用。”

Python:同样的动作,一样的气质

1
2
3
4
5
6
7
8
9
10
11
12
13
from supermemory import Supermemory

client = Supermemory()

client.add(
content="User loves TypeScript and prefers functional patterns",
container_tag="user_123"
)

result = client.profile(container_tag="user_123", q="programming style")

print(result.profile.static) # Long-term facts
print(result.profile.dynamic) # Recent context

它还准备了“即插即用”的框架集成

如果你不想从零接 API,它也准备了 drop-in wrappers(README 里列的生态很全):

  • Vercel AI SDK
  • LangChain
  • LangGraph
  • OpenAI Agents SDK
  • Mastra
  • Agno
  • Claude Memory Tool
  • n8n

示例:Vercel AI SDK / Mastra(README 原样给了代码)

1
2
3
4
5
6
7
// Vercel AI SDK,
import { withSupermemory } from "@supermemory/tools/ai-sdk";
const model = withSupermemory(openai("gpt-4o"), "user_123");

// Mastra
import { withSupermemory } from "@supermemory/tools/mastra";
const agent = new Agent(withSupermemory(config, "user-123", { mode: "full" }));

它像是在说:
“你在用谁都行,我能跟得上。”


搜索模式:它会挑合适的方式“把你需要的东西端上来”

它支持 searchMode

  • hybrid(默认):RAG + Memory 一次给你
  • memories:只在记忆里找(更偏个人化)

README 的示例很直观:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
// Hybrid (default) — RAG + Memory in one query
const results = await client.search.memories({
q: "how do I deploy?",
containerTag: "user_123",
searchMode: "hybrid",
});
// Returns deployment docs (RAG) + user's deploy preferences (Memory)

// Memories only
const results = await client.search.memories({
q: "user preferences",
containerTag: "user_123",
searchMode: "memories",
});

用户画像:一次调用,AI 就知道“自己在跟谁说话”

它把用户画像分成:

  • profile.static:长期稳定事实
  • profile.dynamic:近期动态上下文
1
2
3
4
const { profile } = await client.profile({ containerTag: "user_123" });

// profile.static -> ["Senior engineer at Acme", "Prefers dark mode", "Uses Vim"]
// profile.dynamic -> ["Working on auth migration", "Debugging rate limits"]

这段设计很“像人”:
你不会每次见朋友都背一遍简历,但你也不会忘掉他是谁;
你会在“长期印象”和“最近状态”之间自然切换。
Supermemory 把这种自然做成了 API。


Connectors:让你的知识库不再靠“手动复制粘贴续命”

它支持把外部数据同步进来:

Google Drive · Gmail · Notion · OneDrive · GitHub · Web Crawler

并强调:

  • 实时 webhook
  • 文档自动处理
  • 自动 chunk
  • 可搜索

它像在说:
“你别再当搬运工了,把接口插上,我自己去搬。”


Benchmarks:它不只自夸,还拿成绩单说话

README 写得很硬核:Supermemory 在多个 AI memory benchmark 上表现 state-of-the-art,并且标注:

  • LongMemEval:81.6% — #1
  • LoCoMo:#1
  • ConvoMem:#1

它还做了一个开源框架 MemoryBench,用于标准化评测,并且给了命令行示例:

1
bun run src/index.ts run -p supermemory -b longmemeval -j gpt-4o -r my-run

甚至还准备了一个“让公司评测自己方案”的 Agent skill:

1
npx skills add supermemoryai/memorybench

记忆 vs RAG:它把这句话写得很清楚

它强调了一件经常被混淆的事:Memory 不是 RAG。

  • RAG:检索文档 chunk,偏“无状态”,对每个人一样
  • Memory:抽取并维护“关于用户的事实”,会随时间更新

它举了一个很“人类”的例子:

  • “我刚搬到 SF” 应该覆盖 “我住在 NYC”

并且它默认两者一起跑:你既拿到知识库,也拿到个性化上下文。


自动遗忘:它不会把噪音当永恒真理

它特别写了“Automatic forgetting”:

临时事实会过期,比如:

  • “我明天有考试”

日期过去了,就不该继续占据你的画像。
矛盾会自动解决,噪音不会变成永久记忆。

它像一个很成熟的成年人:
“我记得你,但我不会死记硬背所有废话。”


Links:它给你所有入口(它很开门见山)

��后它又收回那句有点像告白的话:

Give your AI a memory. It’s about time.