Home / AI Build Ideas / P037

AI 開發記憶與規格中樞:讓 Claude Code/Cursor 團隊每次開新 session 都不失憶、需求不跑偏

把每次 agent session 的上下文自動壓縮成長期記憶,並把口頭需求即時整理成正式規格,讓任何相容 client 接上就能『記得』專案。

Difficulty: medium 2–3 週可做出 MVP:先做本機 SQLite 記憶庫+規格轉換單機版,MCP server 包裝與多 client 相容測試再約 1 週;難點在記憶壓縮的取捨(留什麼丟什麼才不失真)與口頭需求轉規格的準確度。

Monetization

①Mac 本機工具買斷(個人開發者版:本機記憶庫+規格產生器,資料不上雲)②團隊版 SaaS 月費(共享記憶庫+規格庫,依席次計費,主打新創/外包 agency)③IDE/編輯器市集擴充功能上架抽成,另接 Ko-fi 贊助與工具/主機商聯盟(非 VPN)。行銷頁廣告只放網頁端 Adsterra/Monetag。

How it works

claude-mem 攔截並壓縮每次 Claude Code/Cursor session 的對話與決策成持久記憶,跨會話載回省去重複解釋專案背景 → sudokar/openspec-plus 把記憶中浮現的口頭需求即時轉成結構化 Spec-Driven 規格文件,做需求發現與驗收條件管理,避免代理程式自己腦補導致需求跑偏 → DietrichGebert/ponytail 讓 agent 依懶惰開發者思維執行,搭配既有記憶與規格減少不必要的重複程式碼與探索步驟 → 全部透過 modelcontextprotocol/python-sdk 包成一個 MCP server,任何相容 client(Claude、Cursor、Windsurf、Copilot)掛上即可共用同一套記憶與規格層。與 P007(Skill 策展)、P021/P027(省 token/理解程式碼)不同:本案核心賣點是『跨會話記憶』與『需求規格化』,解決的是失憶與需求跑偏,不是技能管理或程式碼探索。

Open-source parts

thedotmack/claude-mem

Persistent Context Across Sessions for Every Agent – Captures everything your agent does during sessions, compresses it with AI, and injects relevant context back into future sessions. Works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode + More

🧩 Agent Skills ClaudeCodexGemini ✓ Commercial OK ★ 88,002

sudokar/openspec-plus

OpenSpec Plus — Agentic skills that enhance OpenSpec's Spec-Driven Development through better discovery, requirements, design decisions, execution planning and execution. Works with Claude Code, OpenCode, Github Copilot and any other AI coding agents

🧩 Agent Skills ClaudeWindsurfCopilot ✓ Commercial OK ★ 75

modelcontextprotocol/python-sdk

The official Python SDK for Model Context Protocol servers and clients

🔌 MCP Servers 🔌 Client✦ Curated ✓ Commercial OK ★ 23,670

DietrichGebert/ponytail

Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.

🧩 Agent Skills ClaudeCursor ✓ Commercial OK ★ 86,635