讓 Claude 擁有「長期記憶」的小工具:把你的專案進度存進 Notion,換一個新對話也能接著上次繼續。
A little tool that gives Claude a "long-term memory": it saves your project progress into Notion, so even a brand-new chat can pick up right where you left off.
每次和 Claude 開新對話,它就像失憶一樣,完全不記得你們上次聊了什麼。這個工具就是來解決這個困擾——它把每個專案的進度、決策和踩過的坑記在你自己的 Notion 裡,需要時一個指令讀回來,讓 Claude 好像「記得住」一樣。
Every time you start a new chat with Claude, it forgets everything you discussed before — like it has amnesia. This tool fixes that: it stores each project's progress, decisions, and pitfalls in your own Notion, and reloads them with a single command, so Claude behaves as if it truly "remembers".
市面上讓 AI「記住東西」的工具大多走同一條路:在背景自動偷偷記錄、用 AI 把內容壓縮成摘要、存進一個你看不到的資料庫。這個專案刻意反過來走,核心理念是:你的專案記憶應該可控、可驗證、不失真。
Most tools that give AI a "memory" take the same route: silently record in the background, compress everything into AI-generated summaries, and store it in a database you never see. This project deliberately goes the opposite way. The core idea: your project memory should be controllable, verifiable, and lossless.
① 每一筆記錄都經過你的確認,不會偷偷寫入
只有你下指令(/load、/save)它才會動,存檔前還會先列出「這次要存什麼」讓你確認。不會有 AI 在背景自作主張、寫進一堆你不知道的東西。
① Every record is confirmed by you — nothing is written behind your back
It only acts when you give a command (/load, /save), and before saving it lists exactly what will be stored for your confirmation. No AI silently deciding things and writing who-knows-what in the background.
② 一字不刪的完整存證,不做失真壓縮 很多同類工具靠 AI 自動摘要來省空間,但摘要一定會丟細節。這裡的原則是「逐字保存」:對話中產出的程式碼、prompt、文件,原封不動存進 Notion,存完還會自動核對內容有沒有被截斷或精簡,發現不完整就重存。三個月後回來看,拿到的是完整原文,不是二手轉述。
② Verbatim records, no lossy compression Many similar tools rely on AI auto-summaries to save space — but summaries always lose detail. Here the principle is "store it word for word": code, prompts, and documents produced in a chat go into Notion untouched, and after each save the content is re-checked for truncation and re-saved if incomplete. Come back three months later and you get the full original, not a second-hand retelling.
③ 踩過的坑不會被埋掉
一般的記憶工具只記「發生了什麼」;這個工具把「哪裡出過錯」當成頭等大事。每次載入專案,歷史上所有卡關記錄會被單獨整理成一區「
③ Past pitfalls never get buried
Typical memory tools only record "what happened"; this tool treats "what went wrong" as a first-class citizen. Every time you load a project, all past blockers are pulled into a dedicated "
④ 多個對話同時改,不會互相蓋掉 如果你開了兩個對話串同時處理一個專案,或自己在 Notion 手動改了內容,存檔時工具會用 Notion 的系統時間戳自動偵測「有人比你先改過」,並警告你先重新載入,避免新資料被舊資料無聲覆蓋。這是多數同類工具沒有處理的問題。
④ Multiple chats editing at once won't clobber each other If you have two chats working on the same project, or you edit the Notion page by hand, the tool uses Notion's system timestamp to detect "someone changed this after you loaded it" and warns you to reload first — preventing new data from being silently overwritten by stale data. Most similar tools don't handle this at all.
⑤ 完整的專案生命週期,不只是筆記本
從隨手記下一個點子(/idea),到把點子升級成正式專案(/upgrade idea),到日常進度記錄、一鍵總覽所有專案(/status),再到結案封存(/archive)——它管理的是「點子 → 專案 → 封存」的完整旅程,不只是零散的記憶片段。
⑤ A full project lifecycle, not just a notepad
From jotting down an idea (/idea), to promoting it into a real project (/upgrade idea), to daily progress logs, a one-command overview of every project (/status), and final archiving (/archive) — it manages the whole journey from "idea → project → archive", not just scattered memory fragments.
⑥ 資料放在你自己的 Notion,人看得懂、隨處可用 所有記錄都存在你自己的 Notion 資料庫,不是某個看不到內容的黑盒子。你可以直接打開 Notion 閱讀、編輯、分享給隊友;不管在網頁版 claude.ai、Cowork 還是終端機的 Claude Code,共用的都是同一套記憶。
⑥ Your data lives in your own Notion — human-readable, available everywhere Everything is stored in your own Notion databases, not in an opaque black box. You can open Notion to read, edit, or share it with teammates; and whether you're on claude.ai, Cowork, or the Claude Code terminal, they all share the same memory.
⑦ 針對 Notion 的地雷,內建成套防護 Notion 的 AI 介面有不少隱藏陷阱(例如標籤欄位一寫入就整組覆蓋、表格頁面局部修改會失效)。這個工具把這些坑一個個踩過並寫成了防護規則:重寫整頁前強制先備份、寫入後必定回讀驗證、標籤先讀舊值合併再寫入、超長內容自動分段。你不需要知道這些細節,它自動幫你避開。
⑦ Built-in defenses against Notion's hidden traps Notion's AI interface has quite a few hidden landmines (e.g. tag fields get wholesale-overwritten on write; partial edits fail on pages with tables). This tool has stepped on each of those and turned them into protection rules: mandatory backup before rewriting a page, read-back verification after every write, merging old tag values before writing, auto-splitting oversized content. You don't need to know any of this — it sidesteps the traps for you.
💡 誠實說明取捨:這個工具走「手動、逐字、可驗證」路線,代價是它不會自動幫你存(忘了
/save就沒記錄),也不做模糊語意搜尋。如果你想要全自動、零操作的記憶引擎,別的工具可能更適合;如果你在意每一筆記錄的正確與完整,這就是為你設計的。💡 Honest trade-offs: this tool takes the "manual, verbatim, verifiable" route. The cost: it won't save automatically (forget
/saveand nothing is recorded), and it doesn't do fuzzy semantic search. If you want a fully automatic, zero-effort memory engine, other tools may suit you better; if you care about every record being correct and complete, this was built for you.
以下是最近這幾次的重點更新(最新在最上面):
Here are the most recent highlights (newest first):
- ⚡ 加快存檔速度:在不省略任何安全檢查的前提下,把互不相關的寫入改成同時送出、合併重複的讀取,讓
/save明顯變快。 ⚡ Faster saving: independent writes are now sent in parallel and duplicate reads are merged —/saveis noticeably faster, without skipping a single safety check. - 🛡️ 大幅強化穩定性:針對「資料被舊版覆蓋」「多個對話同時修改互相打架」「長對話後記憶殘缺」等失誤模式做了全面補強。 🛡️ Much stronger reliability: comprehensive hardening against failure modes like "old data overwriting new", "two chats editing at once and clashing", and "degraded memory after long conversations".
- 📄 加入 MIT 授權:專案正式採用 MIT License,任何人都能自由使用。 📄 Added MIT License: the project is now under the MIT License — anyone can use it freely.
- 🔒 機敏資訊分離:把你個人的 Notion 帳號資訊抽到本機專用檔,確保公開專案時不會外洩。 🔒 Separated private data: your personal Notion info now lives in a local-only file, so nothing leaks when the project goes public.
只要三步就能開始使用:
You can get started in just three steps:
第 1 步:下載這個工具 把專案下載下來,放進 Claude 的 skills 資料夾。
Step 1: Download the tool Clone the project into your Claude skills folder.
git clone https://github.com/kau10082/Project_Memory.git projectmemory第 2 步:建立你的設定檔
複製範本,改名成 config.local.json,然後填入你自己的 Notion 資料。
Step 2: Create your config file
Copy the template, rename it to config.local.json, then fill in your own Notion info.
cd projectmemory
cp config.example.json config.local.json第 3 步:重新載入 skill 回到 Claude 重新載入 skill,就可以開始下指令了。
Step 3: Reload the skill Reload the skill in Claude, and you're ready to start giving commands.
💡 設定檔裡要填 4 個 Notion ID:三個資料庫(開頭是
collection://)和一個資料夾頁面(開頭是page://),在 Notion 對應頁面的網址裡就找得到。這個檔案已被.gitignore排除、不會被上傳,請放心填。💡 The config needs 4 Notion IDs — three databases (starting with
collection://) and one folder page (starting withpage://); you can find them in the URL of each Notion page. This file is excluded by.gitignoreand never uploaded, so it's safe to fill in.
先講一個重點:在不同環境呼叫指令的方式不一樣。
One key note first: how you call the commands depends on your environment.
- 在網頁版(claude.ai / Cowork):直接打斜線指令,例如
/load EBM。 On the web (claude.ai / Cowork): type slash commands directly, e.g./load EBM. - 在終端機(Claude Code):改用「projectmemory + 動作」的說法,例如「用 projectmemory skill load EBM」,因為純斜線會和終端機內建指令打架。 In the terminal (Claude Code): say "projectmemory + action" instead, e.g. "use the projectmemory skill to load EBM", because bare slashes clash with the terminal's built-in commands.
最常用的流程就是:開新對話 → /load 讀回記憶 → 做事 → /save 存回去。常用指令如下:
The most common flow is: new chat → /load to recall → work → /save to store back. Here are the main commands:
| 指令 / Command | 白話說明 / Plain-language meaning |
|---|---|
/load [專案] |
讀回某個專案的記憶,連同歷史踩坑記錄一起帶回來。 Reload a project's memory, pitfall history included. |
/save [專案] |
把目前進度存回 Notion(會先列清單讓你確認)。 Save current progress back to Notion (lists everything for confirmation first). |
/save all |
一次存下這串對話裡討論過的所有專案。 Save every project discussed in this chat at once. |
/idea [想法] |
隨手記下一個新點子到點子池。 Jot down a new idea into the idea pool. |
/upgrade idea [點子] to [專案] |
把成熟的點子升級成正式專案。 Promote a ripe idea into a real project. |
/status |
看所有進行中專案的狀態總表。 See a status table of all active projects. |
/archive [專案] |
把完成的專案標記為已封存。 Mark a finished project as archived. |
專案的檔案很單純,結構如下:
The project has a very simple structure:
Project_Memory/
├── SKILL.md # 工具本體:所有指令的定義、流程與防護規則
├── config.example.json # 設定範本(給你複製用)
├── config.local.json # 你的真實 Notion 資料(本機才有,不會上傳)
├── README.md # 說明文件(就是你正在看的這份)
├── LICENSE # 授權條款
└── .gitignore # 指定哪些檔案不要上傳
各檔案的作用(白話版):
What each file does (in plain words):
SKILL.md:整個工具的大腦,寫著每個指令該怎麼運作、以及所有資料安全防護規則。SKILL.md: the brain of the tool — it defines how each command works, plus all the data-safety protection rules.config.example.json:一份空白範本,讓你照著填自己的設定。config.example.json: a blank template for you to copy and fill in.config.local.json:填了你真實 Notion 資料的檔案,只存在你電腦裡、不會上傳,是隱私安全的關鍵。config.local.json: the file with your real Notion info — it stays only on your computer and is never uploaded; this is the key to your privacy.README.md/LICENSE/.gitignore:分別是說明書、授權條款、以及「不要上傳清單」。README.md/LICENSE/.gitignore: the manual, the license, and the "do-not-upload list" respectively.
非常歡迎你一起讓這個專案變得更好!不論是修正錯字、改善功能,或只是回報一個小問題,都幫助很大。
You're very welcome to help make this project better! Whether it's fixing a typo, improving a feature, or just reporting a small issue, every bit helps.
想回報問題? 到專案的 Issues 頁面開一則新的 issue,把你遇到的狀況描述清楚就好,不用客氣。
Want to report a problem? Just open a new issue on the project's Issues page and describe what you ran into — don't be shy.
想動手改程式? 可以照這個流程:
Want to change the code yourself? You can follow these steps:
- 先把專案 Fork(複製)到你自己的帳號。 Fork (copy) the project to your own account first.
- 建立一個新的分支來做你的修改。 Create a new branch for your changes.
- 改好後發一個 Pull Request(PR),說明你改了什麼。 Open a Pull Request (PR) describing what you changed.
- 我們會盡快看過並回覆你。 We'll review it and get back to you as soon as we can.
本專案採用 MIT 授權條款。簡單說:你可以自由地使用、修改、散布,甚至用於商業用途,唯一的條件是保留原作者的版權聲明。本專案不負任何擔保責任。
This project is under the MIT License. In short: you're free to use, modify, distribute, and even use it commercially — the only condition is that you keep the original copyright notice. The project comes with no warranty.
完整條款請見 LICENSE 檔案。
See the LICENSE file for the full terms.