Minnimemory

io.github.MinniAI-com由 GitHub 帳號 MinniAI-com 發布v1.4.2更新於 Nov 7, 2026

Curated, on-demand memory for AI coding agents. Cuts Claude Code token cost per session.

已驗證STDIO僅桌面知識與記憶開發者工具

概覽

AI 產生的概覽

本機 MCP 伺服器,稽核 Claude Code 專案的記憶檔案,並編譯成精簡的常駐記憶與按需載入的記憶片段。

可以做什麼
  • 稽核並列出 Claude Code 在專案中載入的記憶檔案
  • 預覽編譯方案並執行,回傳需使用者核准的指令
  • 檢查上次執行後的漂移並產生重新執行方案
  • 用 undo 逐位元組還原原始記憶檔案
  • 透過鉤子在提示中附加相符的記憶片段,略過疑似憑證的行
適用情境
當 Claude Code 專案的 CLAUDE.md、AGENTS.md 或記憶資料夾變得臃腫,想降低每輪工作階段的 token 成本又保留可還原的原始檔案時適用。也適合定期檢查記憶檔案是否發生漂移。
執行需求
以 Claude Code 外掛(或 npm 套件 minnimemory)形式透過 stdio 在本機執行。需要 Claude Code 看得到的 PATH 中有 Node.js 18.3 或更新版本,一鍵安裝外掛還需 Claude Code 2.1.275 或更新版本。僅限桌面端,在 claude.ai 網頁對話中不生效。未宣告任何 API 金鑰或環境變數。
安裝前請注意
apply 和 undo 會改動磁碟上的記憶檔案,但會先備份原始檔案,undo 可還原。它會在 Claude 設定資料夾和專案的 .minnimemory 目錄寫入紀錄。同時安裝外掛和 npx 方式會導致伺服器重複註冊。token 節省數據僅來自廠商自己的記憶資料夾。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Minnimemory,將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

MinniMemoryMCP for Claude Code

MinniMemoryMCP 1.4.2 as a Claude Code plugin. It audits the memory files Claude Code loads (CLAUDE.md, AGENTS.md, your memory folder) and compiles them into a small AlwaysOnMemory body that loads every turn, plus OnDemandMemory files that load only when a task needs them. Every change is lossless and undoable: the originals are backed up first and undo puts them back byte for byte. Everything runs on your machine, on your own files, only when you ask. Built by MinniAI.

Requirements

  • Node.js 18.3 or later on your PATH. The server and the hooks run with node. Claude Code's own installer does not need Node, so check with node --version first.
  • Claude Code 2.1.275 or later for the one-line install below; older versions add the marketplace first (the shell form under it).

Install

/plugin install minnimemory --marketplace MinniAI-com/minnimemory-plugin

Or from your shell: claude plugin marketplace add MinniAI-com/minnimemory-plugin, then claude plugin install minnimemory@minniai.

The plugin loads in Claude Code, and in Cowork when the session runs on your computer. In claude.ai chat it is ignored: a local MCP server and hooks do not run there.

Pick one route per machine. If you ran npx minnimemory install before, run npx -y minnimemory uninstall first: otherwise the server is registered twice and every prompt starts two copies of each hook (the second copy stays silent, but you pay its start).

Use

Three prompts cover the whole loop, in your project in Claude Code:

  1. "Clean up my memory with minnimemory." Claude calls init, which shows the map of what Claude Code loads here, the audit and the change apply would make. It writes nothing but a small record of where your memory loads from.
  2. "Apply it." Claude calls apply, which hands back one command and the list of file operations; the command runs only through your approval prompt.
  3. A week later, "check my memory." Claude calls doctor, which shows what drifted and the re-apply plan; "undo the memory compile" restores the originals.

The same four steps are slash commands: /mcp__minnimemory__init, doctor, apply, undo.

What it runs, reads, writes and sends

  • An MCP server (node server/cli.mjs mcp) with four tools: init, doctor, apply, undo. It reads only the folder a tool call names and the files you or Claude tell it Claude Code loads; nothing is searched for. apply and undo return a command and the list of file operations, and Claude runs that command only with your approval. The command is node "<plugin folder>/server/cli.mjs" apply --plan <hash> <folder>: the plugin's own file, never a download.
  • Three hooks. Before each prompt, drift checks whether your memory changed since the last apply and prints one line when it did; hook adds the one short section of memory that matches the prompt, inside a block marked as data, once per session, and never a section that holds a credential-shaped line (it says so instead). The prompt hook's saving was measured in live Claude Code sessions on our own memory folder (the number is on the docs page); on other folders it is unmeasured. When Claude Code loads its instructions, loaded records which file it loaded in a local log that init can read later.
  • Also in the bundle, never run by the plugin. server/cli.mjs is the npm package's command line, so it also contains the install and uninstall commands of the npx route, which run claude mcp add and claude mcp remove and write the npx form of the hooks into your settings, and a mcp --profile full maintainer mode with six more tools (recall, outline, list, audit, plan, compile). The plugin starts the server with mcp alone, the four-tool profile above; nothing in the plugin calls the rest. They are there for someone who runs the file by hand.
  • What "memory" means here. Your own files on disk: CLAUDE.md, the files it imports, and the memory folder Claude Code keeps for a project, read only when you point a tool at them. The plugin never reads chat history, conversation summaries or uploaded files.

Try it on a sample first

examples/CLAUDE.md in this folder is a fictional project's memory file. Copy it into an empty folder, open Claude Code there and say "Clean up my memory with minnimemory", then "Apply it", then "undo the memory compile". That is the whole loop on a file you cannot break.

Security reports

Report a vulnerability to [email protected] or as a GitHub issue on this repository; both are read by the maintainer. The plugin makes no network calls, so most reports will be about what it reads or writes on disk.

  • What it writes, and where. Under your Claude config folder (~/.claude unless CLAUDE_CONFIG_DIR says otherwise): minnimemory/discovery.json (where your memory loads from, written by init), minnimemory/loads/<project>.jsonl (the load log), minnimemory/hook-sessions/<session>.json (which sections the prompt hook already placed), and minnimemory/undo-backups/ (a copy of what a forced undo discards). In a compiled project or memory folder: .minnimemory/ holds the backups the apply command makes and a small drift stamp. No memory file is ever written by the server itself.
  • Nothing is sent anywhere. The plugin makes no network requests, calls no model and collects no data. The only data it reads from your session is the prompt text the hook matches against your memory, and that is not stored.

Troubleshooting

  • "hook failed to start" on every prompt, or the server does not connect. Node is missing or not on the PATH Claude Code sees. Install Node 18.3 or later; a Claude Code started from a desktop launcher may need a restart after the install so it picks up the new PATH.
  • The prompt hook never adds anything. It places a section only when it clearly matches the prompt, is small enough to be cheaper placed than read, and is not already in the always-loaded body. A tiny or uncompiled memory gives it nothing to place; run init and apply first.
  • Two servers named minnimemory. You have both the npx install and the plugin. Run npx -y minnimemory uninstall and restart Claude Code.
  • apply says the plan hash does not match. A memory file changed since the preview. Run doctor for a fresh plan, then apply again.

Issues: https://github.com/MinniAI-com/minnimemory-plugin/issues. Documentation: https://minniai.com/minnimemorymcp/docs.

Updates

Turn on auto-update for the minniai marketplace in /plugin, or run claude plugin update minnimemory@minniai. A new version loads on the next launch or after /reload-plugins.

Company admins: keep auto-update off and pin the version in managed settings when every machine must run the same release.

Remove

claude plugin uninstall minnimemory@minniai. Compiled memory stays compiled; run undo first if you want the originals back. The small records under <config>/minnimemory/ stay too; delete that folder if you want nothing left.

Licence

See LICENSE. Third-party code bundled in server/ is listed with its licences in THIRD_PARTY_NOTICES.md.

來源:README.md,提交 5f248ef

工具

0
SourceWeft 尚未在沙箱中啟動過該軟體包,因此其工具未知。

版本歷史

3
  1. v1.4.2最新Nov 7, 2026
  2. v1.4.1Oct 29, 2026
  3. v1.3.7Oct 22, 2026

評分

已滿足適用檢查項的 40%

該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通25 分

    請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。

  • 已宣告認證方式10 分

    請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。

  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分

    請在儲存庫或註冊中心條目中宣告授權條款。