ViBo Memory

io.github.vnbochkarev-netizen由 GitHub 帳號 vnbochkarev-netizen 發布v0.3.0更新於 Nov 8, 2026

Local-first memory for AI agents: facts with hash-chain provenance, verified recall.

發布方自己的服務:Memory.fit

已驗證STDIO僅桌面知識與記憶AI 與機器學習

概覽

AI 產生的概覽

為 AI 助理設計的本地優先記憶伺服器:把事實存到本機 JSON 檔,可搜尋、核對說法並稽核雜湊鏈。

可以做什麼
  • 把長期事實存進本機記憶,依標籤與內容去重,並區分 L1/L2/L3 機密等級
  • 依關鍵字搜尋已存事實並依相關度排序,除非明確要求否則隱藏機密等級事實
  • 對照記憶核對說法,回傳已確認、矛盾、部分符合或找不到
  • 稽核記憶:重算雜湊鏈並列出事實的來源與時間戳
  • 依精確標籤或 id 刪除事實,需明確確認
  • 檢視記憶規模與節省的 token,並維護對話記憶:加入訊息、讀取上下文、搜尋或把較舊訊息壓縮成摘要
適用情境
適合希望助理跨工作階段記住關於你、專案、數字或日期的穩定事實、又不想使用雲端帳號的情境,也適合在重述事實前先對照記憶核對。需要追溯某條事實來源時同樣適用。
執行需求
以本機 stdio 程序執行,從 npm 套件安裝,需要 Node.js 與 MCP 用戶端。預設不需帳號、金鑰或網路請求。選用的 VIBO_API_KEY 會把事實鏡像到 ViBo 雲端帳號,VIBO_MEM_FILE 可覆寫預設記憶檔位置。
安裝前請注意
記憶寫入本機 JSON 檔,儲存的內容會留在磁碟上,能存取該檔案的人都能讀取。刪除事實無法復原,且需要 confirm=true。設定 VIBO_API_KEY 會把事實傳送到第三方雲端服務。標記為機密等級的事實預設不會出現在搜尋結果中,但仍保存在本機。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 ViBo Memory,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

ViBo MCP Server

Local-first memory for AI agents. Facts live on your machine in one JSON file, each one chained to the previous by hash, so you can prove later where a fact came from and that nobody edited it. Works with any MCP client: Claude Desktop, Claude Code, Cursor, Codex, VS Code, Windsurf.

No account, no key, no network call required.

Install

bash
npm install -g @vibo-dev/vibo-mcp

Client config (Claude Desktop example):

json
{  "mcpServers": {    "vibo": {      "command": "npx",      "args": ["-y", "@vibo-dev/vibo-mcp"]    }  }}

Memory is stored in ~/.vibo/mcp/memory.json. Override with VIBO_MEM_FILE (file path) or VIBO_HOME (directory).

Tools

ToolWhat it does
memory_addSave a fact. Deduplicated by label + content. Levels L1 / L2 / L3 (secret).
memory_searchFind facts by words, ranks by relevance. L3 hidden unless include_secrets.
memory_verifyCheck claims: CONFIRMED / CONTRADICTS / PARTIAL / NOT FOUND. Catches wrong numbers.
memory_provenanceVerify the hash chain and list facts with origin and timestamps.
memory_forgetDelete by label or id. Refuses without confirm=true.
memory_usageFacts, tokens stored, average tokens returned per search.
thread_memoryConversation memory: add, context, ask, compress older messages.
web_searchNot in this build — returns an honest "unavailable" instead of failing silently.

Example

memory_add(label="client-anna-prefs", content="Anna orders coffee without sugar, budget 24000 RUB per month")memory_search(query="what does Anna prefer")  -> [L1] client-anna-prefs: Anna orders coffee without sugar, budget 24000 RUB per month (0.86)
memory_verify(claims=["Anna budget is 24000 RUB per month", "Anna budget is 99000 RUB per month"])  -> CONFIRMED "Anna budget is 24000 RUB per month" — memory: client-anna-prefs (L1)  -> CONTRADICTS "Anna budget is 99000 RUB per month" — memory: client-anna-prefs (L1)

Design

  • Local only by default. The server never phones home. If you set VIBO_API_KEY and VIBO_BASE_URL, facts are mirrored best-effort; failure never breaks a tool call.
  • Provenance built in. Every fact carries hash and prev; memory_provenance recomputes the chain and reports the first break.
  • Secrets stay put. L3 facts are excluded from search results unless explicitly requested.
  • Deterministic recall. Lexical ranking with light stemming, no embeddings, no background "consolidation" that silently rewrites your memory.

Tests

bash
npm run buildpython3 tests/smoke.py    # 20 checks over the real stdio protocol

Links

Integrations

來源:README.md,提交 98b6781

工具

8

memory_add

Save a durable fact to local memory on this machine. Deduplicated by label+content.

輸入參數label必填levelcontent必填

memory_verify

Check factual claims against local memory: CONFIRMED, CONTRADICTS, PARTIAL or NOT FOUND per claim. Use before stating facts about the user, projects, numbers or dates.

輸入參數modeclaims必填

memory_provenance

Audit local memory: verify the hash chain, show facts with origin and timestamps.

輸入參數limit

memory_forget

Delete a fact by exact label or id. Requires confirm=true.

輸入參數target必填confirm

memory_usage

Show local memory size and how much context it saves versus pasting facts into the prompt.

無需輸入參數。

thread_memory

Thread memory: add a message, read recent context, search the thread, or compress older messages into a digest.

輸入參數keeproletextquerytopicaction必填

版本歷史

2
  1. v0.3.0最新Nov 7, 2026
  2. v0.2.6Sep 16, 2026

評分

已滿足適用檢查項的 80%

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
  • 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。

  • 已列出工具必需15 分
  • 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。

其他 Memory 整合