Feedback Memory

MCP文档知识

提供方 未标注

文档知识 stdio · 本地进程 未标注 社区 / 第三方 Claude DesktopClaudeCursorVS Code
查看接入方式

记住你批准、编辑或拒绝智能体草稿的原因,并在下一次草稿前添加最接近的历史修正。MCP服务器 + Claude Code插件,本地Postgres + pgvector。

一句话选型结论

按资料字段整理

Feedback Memory 属于文档知识类 MCP 服务器,stdio · 本地进程,鉴权方式未标注,可直接接入 Claude Desktop、Claude、Cursor 等,有可直接执行的安装命令。接入前先确认权限范围;涉及写入、删除或付款的操作,建议先在测试环境验证。

分类文档知识 传输方式stdio · 本地进程 鉴权未标注 部署方式本地部署
GitHub 星标 未标注 仓库热度,无公开仓库时显示未标注
工具数 未标注 暴露给模型的工具数量
资源 / 提示词 0 / 0 资源与提示词模板数量
维护方 社区 / 第三方 核验等级 仅资料核对

工具与能力

文档知识
数量来自来源目录的标注,具体工具清单以官方文档为准
工具数量未标注暴露给模型可调用的工具个数
资源数量0可读取的资源条目
提示词数量0内置提示词模板数量

能力标签

支持客户端

适用场景

记忆

安装与配置

stdio · 本地进程 未标注
客户端在本机拉起子进程,数据不出本机
claude mcp add feedback-memory -- npx -y github:ssap-pa/self-learning-agent-setup

这条命令只是入口,具体启动参数请按官方文档补齐,不要直接照搬到别的服务器。

  1. 1
    准备运行环境

    stdio 服务在本地起进程,需要先装好 Node.js 等运行时,再按安装命令拉取依赖。

  2. 2
    安装服务

    复制上面的安装命令执行;命令里带占位符的位置,换成自己的密钥或路径。

  3. 3
    写入客户端配置

    在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,让新配置生效。

  4. 4
    验证是否可用

    连接成功后随便调一个只读工具,确认能返回结果,再接入正式流程。

典型使用场景

3 类
按分类与标签整理的常见用法,具体以官方文档为准
文档检索问答

在你的文档库里按问题找答案。

知识库读写

把新资料写进知识库,随用随取。

内容整理

把零散资料归纳成结构化笔记。

选择建议

选之前先看这几条
按条目字段给出的判断,不替你决定用不用

值得选的地方

  • 有现成的安装命令,不用自己找入口,照着跑就能起服务。
  • 已适配 Claude Desktop、Claude、Cursor,换客户端不用重新研究配置。

先想清楚的地方

  • stdio 方式要求本地装好运行时(如 Node.js),换机器要重新装一遍。
  • 资料没标注工具数量,实际能力要装好后确认。
  • 版本更新与行为细节以官方仓库为准。

更适合

  • 需要在文档知识场景里让模型直接调用外部能力的人。
  • 已经在用 Claude Desktop、Claude 的用户,接入路径最短。
  • 想先快速试一下、能接受在本地或自有环境跑服务的人。

不太适合

  • 完全不想在本地装环境的人。
  • 需要厂商级 SLA 和技术支持的生产环境。

同类对比

6 个同类
同分类里挑最接近的几条,字段以各自详情页为准
MCP 服务器工具数传输方式部署方式安装命令最后核验
Feedback Memory当前 未标注 stdio · 本地进程 本地部署 有 未标注
dify 未标注 stdio · 本地进程 本地部署 有 2026-04-11
open-webui 未标注 stdio · 本地进程 本地部署 有 2026-04-11
JeecgBoot 未标注 stdio · 本地进程 本地部署 有 2026-04-11
Cherry Studio 未标注 未标注 未标注 无 2026-05-17
mempalace 未标注 stdio · 本地进程 本地部署 有 2026-04-11

服务器档案

文档知识
字段来自项目仓库与公开索引,中文说明由本站标注
提供方未标注上游标注的维护方 未标注
分类文档知识本站按上游分类展示
连接方式stdio · 本地进程客户端在本机拉起子进程,数据不出本机
鉴权方式未标注远程服务建议用最小权限凭据
维护方社区 / 第三方核验等级 仅资料核对
包名 / 运行时未标注 · 未标注本地安装时用到的包标识与运行时
部署类型本地部署上游给出的部署形态
最近核验未标注上游最后核验时间

详细介绍

9.4 KB
来自项目仓库与上游目录的原文,未做删改

Self-Learning Agent Setup

A small, practical pattern for making an AI agent remember your business and learn from every correction you give it. No framework required: a folder of brand files, your existing database, and a feedback table.

I run a one-person education business in Korea this way. Agents (Claude Code, Codex) draft posts and replies. I approve, edit or reject them from my phone. Every decision goes into a database, and the next draft reads that history before it writes a word.

This repo has the minimal schema and the prompts. The full walkthrough is a free 23-page field guide (link at the bottom).

> What happened when this agent ran a business for 72 hours with one goal, in numbers: the postmortem on DEV. Short version: product was never the problem.

Try it in one minute

In your browser, no install: https://ssap-pa.github.io/self-learning-agent-setup/ (Postgres + pgvector via PGlite and the all-MiniLM-L6-v2 embedding model via Transformers.js, all running in the tab, no API key).

Locally:```
git clone https://github.com/ssap-pa/self-learning-agent-setup
cd self-learning-agent-setup/demo
npm install
npm run demo


It starts an in-memory Postgres (PGlite + pgvector), feeds in three days of feedback for a made-up candle shop (one edit, one reject, one plain approval), then prints what the agent would read before its next drafts. No database or API key needed. `npm test` runs the tests the same way.```
Day 3  new comment: "How long does shipping to Canada take for the lavender set?"
       before drafting, the agent reads:
         Feedback on similar past tasks (most similar first):
         - task: "Do you ship to Canada? Love the lavender candle" -> edited to: "We do! Canada usually takes 5-7 days." (why: don't promise speed, give the real shipping time)

Day 3  new comment: "Could you gift wrap a candle for a birthday?"
       before drafting, the agent reads:
         Feedback on similar past tasks (most similar first):
         - task: "Can you gift wrap the vanilla set for my mom?" -> rejected (why: we don't gift wrap, offer the gift note instead)

Day 3  new comment: "What wax do you use?"
       before drafting, the agent reads:
         (nothing yet: no feedback on anything similar)

The demo embeds text by hashing words, so "similar" here means shared words. Run EMBED=openai OPENAI_API_KEY=... npm run demo to use real embeddings, or point Loop at your own Postgres / Supabase with pg.

Use it in Claude Code, Cursor or Claude Desktop (MCP server)

feedback-memory is a small MCP server built on the same loop. Your agent calls recall_corrections before it drafts and record_decision after you approve, edit or reject. Everything lives in a local Postgres (PGlite + pgvector) in ~/.feedback-memory, so there's nothing to host.

Claude Code, as a plugin (the easiest way). Your standing rules load when a session starts (and again after /clear or compaction), and the closest past corrections are added to each prompt by a hook, so they apply even when Claude doesn't call a tool:```
/plugin marketplace add https://github.com/ssap-pa/self-learning-agent-setup.git
/plugin install feedback-memory@ssap-pa


(The short form `/plugin marketplace add ssap-pa/self-learning-agent-setup` works too if you have an SSH key set up for GitHub.)

In a real run with the plugin, Claude made no tool calls and still answered a new "how long does shipping to Canada take?" comment with "Thanks for asking. Shipping to Canada usually takes 5-7 business days for the lavender set.": the shipping time from a past edit (reason: don't promise speed) and no exclamation marks, from a stored rule.

Claude Code, MCP server only:```
claude mcp add feedback-memory -- npx -y github:ssap-pa/self-learning-agent-setup

Claude Desktop, one click: download feedback-memory.mcpb and open it. The OpenAI key field is optional.

Cursor, or Claude Desktop by hand (under mcpServers in the config JSON):```json
"feedback-memory": { "command": "npx", "args": ["-y", "github:ssap-pa/self-learning-agent-setup"] }


Then tell the agent when to use it, e.g. two lines in `CLAUDE.md` / `AGENTS.md`:```
Before drafting anything I'll review, call recall_corrections with the task and follow what comes back.
When I approve, edit or reject your draft, call record_decision with my reason.

Tools: recall_corrections, record_decision, add_rule (standing rules that always come back first), list_memory, forget. Set OPENAI_API_KEY in the server's env to match by meaning (text-embedding-3-small); without it, words are hashed offline (with light stemming, so "shipping" matches "ship") and only shared words match. One database folder per running client (set FEEDBACK_MEMORY_DIR if you run several), so install it one way, not as both the plugin and a separate MCP server. npm test runs an end-to-end test over stdio, and npm run bundle builds the Claude Desktop bundle.

If it helps, a star on the repo helps other people find it, and an issue about what didn't work helps me fix it.

The problem

Anyone can pay for the same model you use. What nobody else can buy is:

  1. what your agent knows about your business, and
  2. the record of every correction you've given it.

Most people give feedback in chat and lose it. The next session starts from zero, and you fix "stunning blooms" for the tenth time.

Layer 1: the knowledge layer (what it knows)

Type Example Where it lives
Constants Brand, voice, customer, founder story Markdown files in a brand/ folder
Variables Price, stock, seats, schedule Your site's database, read live
Large knowledge Lessons, ebooks, FAQs Same database, embedded for search (RAG)

Rules that matter:

  • Files for who you are, database for what's true right now. Prices in a text file go stale, and the agent quotes last month's promo with confidence.
  • The agent can only say what your admin page stores. "3 seats left" requires a seats field.
  • Enforce it in code. A rule that only lives in AGENTS.md can be skipped. Programs that run on their own should call the database and search directly.

Layer 2: the learning loop (how it improves)

  1. Store every approve (+1), reject (−1) and edit (0, with the correction), linked to the run it judges.
  2. Generate candidates. Repeated feedback becomes a proposed rule, not an active one.
  3. Score with a rubric per goal.
  4. Compare the candidate against current behavior.
  5. Promote or drop.

Plus two details that made it actually work for me:

  • Rules vs. memory. "Your tone sounds robotic" always applies (rule). "For funeral orders, no emojis" only applies when the situation matches (memory). Memories that keep repeating get promoted.
  • Search feedback by meaning. Keyword search missed feedback phrased differently. Embeddings pull in corrections from similar situations.

And the human part: a Telegram bot sends each draft to my phone with Approve / Reject / Edit buttons, and the taps land in the learning database (not just the bot's log). Check that they actually do. Mine silently wasn't storing approvals at first.

What's here

  • schema.sql: Postgres / Supabase tables (runs, feedback with situation and reason embeddings, rules) and match_feedback() for similarity search, with pgvector
  • demo/loop.ts: the core loop in about 100 lines of TypeScript: store each decision, pull feedback from similar past tasks into the next prompt. Works with pg or PGlite
  • demo/demo.ts and demo/loop.test.ts: the one-minute demo and tests
  • prompts.md: the prompts I use with Claude Code / Codex to build and check each layer

Start small

  • Run one goal with low limits (mine: 1 post and up to 3 replies a day for two weeks).
  • Keep approval on until there's nothing left to correct.
  • Measure purchases or signups, not likes.

More

  • Need it across machines or a team, or approvals from your phone? The free plugin keeps its memory in one folder on one machine. The Self-Learning Agent Kit runs the same loop on your own Postgres / Supabase so every machine and teammate shares one memory, adds a Telegram bot to approve / edit / reject from your phone (every tap is stored before anything publishes), proposes rules from reasons that keep coming back, and ships Python and TypeScript versions of the library + CLI, an MCP server and its own Claude Code plugin whose hooks read that shared database (by meaning, with OpenAI embeddings), an end-to-end example and 28 tests ($39): https://payhip.com/b/grc25
  • Prefer video? A free 22-minute lesson where I build this system with an agent (Korean audio, English subtitles, no account): https://ssapable.com/courses/ai-agent?lang=en&utm_source=github&utm_medium=readme#free-preview
  • Free field guide (23 pages, PDF): the full walkthrough with a worked example and a 7-day plan: https://payhip.com/b/QHgfJ
  • The book: Just Say "Do It", an 11-step playbook for running a one-person business with AI agents (English edition of my Korean course): https://payhip.com/b/xmZvu. Pay what you want until Sunday, Oct 4 (US Pacific), and $0 is fine; if you read it, an honest review on that page helps a lot.
  • Want it set up around your business? Done-for-you automation blueprint: https://payhip.com/b/BzSRC

Text and diagrams (this README, prompts.md, img/) © 2026 AI SSAPABLE, shared under CC BY-NC 4.0. The MIT License in LICENSE covers the code: schema.sql, demo/, docs/, mcp/, hooks/, bundle/ and .claude-plugin/.

常见问题

8 问
答案由本站依据条目字段整理,不引用上游原文
Feedback Memory 怎么安装?

在客户端按下面的安装命令接入:claude mcp add feedback-memory -- npx -y github:ssap-pa/self-learning-agent-setup。装完在客户端配置里引用同一个包或端点。

Feedback Memory 怎么在 Claude Desktop、Claude、Cursor 里配置?

它的传输方式是stdio · 本地进程,在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,配置才会生效。

Feedback Memory 收费吗?

本站不代收费用,也没有标注具体价格。是否收费取决于服务本身和背后的接口额度,正式使用前按官方说明确认。

Feedback Memory 需要 API Key 吗?

资料没有标注鉴权方式,是否需要密钥取决于服务本身,接入前按官方文档确认。

Feedback Memory 支持哪些客户端?

资料标注支持 Claude Desktop、Claude、Cursor、VS Code、Codex、Zed;其他客户端只要能配置 MCP Server,也能用同样的命令或地址接入。

Feedback Memory 能提供哪些能力?

资料显示能力标签包括 记忆。具体工具清单以官方文档为准。

Feedback Memory 需要什么运行环境?

属于本地进程,需要在本机装好运行环境。

Feedback Memory 是本地还是远程?

它是本地进程(stdio),由客户端在本机拉起。传输方式为stdio · 本地进程。

热门专题

按标签挑服务器
全部专题 →

服务器信息以作者仓库与官方文档为准;分类、传输方式与鉴权方式的中文说明为中文整理,最近核验 未标注。