AI Unified Memory (AUM)

August 18, 2026 · View on GitHub

One memory warehouse, many AI agents. Public library + private libraries + cross-AI messaging + fully automated scheduler.

一个记忆仓库，多个 AI 共享。公用库 + 专有库 + AI 间消息 + 全自动调度器。

🖼️ 界面预览

AI Unified Memory GUI

图形化客户端：记忆浏览 / 语义检索 / 调度中心 / 跨AI消息 / 热记忆

📥 立即下载（三平台安装包，双击即用）

平台	轻量版（基础管理）	完整版（记忆浏览/语义检索/消息）
🪟 Windows	⬇️ AI-Unified-Memory-Windows.exe (10.6MB)	⬇️ AUM-Windows-x64.exe (17.4MB)
🍎 macOS	⬇️ AI-Unified-Memory-macOS (9.2MB)	⬇️ AUM-macOS.app (14.8MB)
🐧 Linux	⬇️ AI-Unified-Memory-Linux.AppImage (20MB)	⬇️ AUM-Linux-x86_64.AppImage (27.4MB)

💡 轻量版 = tkinter 纯标准库（零依赖，启动最快） · 完整版 = pywebview 现代界面（记忆浏览/语义检索/冲突消解/热记忆/跨AI消息） 🔗 全部版本：Releases 页面

English Introduction

What it is

When you run multiple AI agents (Hermes, Codex, OpenClaw, Qoder, WorkBuddy, Claude Code…), each one keeps its own private memory. The result:

❌ Knowledge silos: what one AI learned is invisible to the others
❌ Repetition: every AI re-learns the same user preferences and project facts
❌ No coordination: no way to pass messages, tasks, or lessons between AIs
❌ No evolution: memory never gets deduplicated, categorized, or versioned

AUM turns scattered per-AI memories into one shared, evolving, self-maintaining memory system:

✅ Public Library — authoritative shared memory (user profile, project knowledge, domain knowledge, lessons learned, decisions), categorized & deduplicated
✅ Private Libraries — per-AI mirrors of raw memories + an auto-generated injection file each AI reads at session start
✅ Exchange Area — INBOX/OUTBOX messaging between AIs (task dispatch, notifications)
✅ Scheduler — one command runs the full loop: scan → promote → dispatch → index → snapshot
✅ Zero third-party dependencies — pure Python standard library, runs anywhere

Key Features

🔄 One-click sync: scan → promote → dispatch → index → snapshot, fully automated
🔍 Cross-library search: full-text keyword search across public + private + exchange
📊 Memory stats: live file statistics for every library
🖥️ Graphical client: no coding required — download an installer, double-click and go
🧠 Auto-maintenance: global dedup, keyword-scored auto-classification, daily snapshots
💬 Cross-AI messaging: send/read messages between AIs via INBOX/OUTBOX
📦 Zero dependencies: pure Python standard library, no pip install needed

Installers (Graphical Client)

No programming required — download, double-click, done:

Platform	Download
🪟 Windows	`AI-Unified-Memory-Windows.exe` (download & run)
🍎 macOS	`AI-Unified-Memory-macOS` (App)
🐧 Linux	`AI-Unified-Memory-Linux.AppImage`

GUI features:

🔄 One-click sync: scan → promote → dispatch → index → snapshot, fully automatic
🔍 Search memory: keyword search across all libraries
📊 Memory library status: per-library file statistics at a glance
⚡ Dark professional theme; every operation is a single click

All the complex sync/classify/dispatch logic runs in the background — the user only needs to click one button.

Developer Mode

python gui_app.py                     # launch the GUI
python scripts/coordinator.py --full  # or full sync from the command line

🖥 Desktop App (Windows / macOS / Linux)

Download the installer for your platform from Releases:

Platform	Installer
Windows	`AUM-Windows-x64.exe` (double-click to run)
macOS	`AUM-macOS.app` (drag to Applications)
Linux	`AUM-Linux-x86_64.AppImage` (chmod +x, run)

Features: memory browser · semantic search (synonym-expanded TF scoring) · one-click scheduler · conflict resolution · cross-AI messaging · hot-memory ranking.

# Build locally (requires Python 3.8+)
pip install -r requirements-desktop.txt
bash scripts/build_client.sh        # auto-detects platform

🚀 Quick Start

# 1. Copy config template and fill in your AI memory source paths
cp CONFIG.example.json CONFIG.json

# 2. Run the full synchronization loop (scan → promote → dispatch → index → snapshot)
python scripts/coordinator.py --full

# 3. Search shared memory on demand
python scripts/search.py "user profile" --limit 10

# 4. Cross-AI messaging
python scripts/msg.py send --to Codex --title "Please review" --body "..."
python scripts/msg.py list
python scripts/msg.py read <message-id>

Architecture

                     ┌───────────────────────────────────┐
                     │         coordinator.py            │
                     │  scan → promote → dispatch → snap │
                     └───────┬───────────────┬───────────┘
                             │               │
              ┌──────────────▼──┐   ┌────────▼───────────┐
              │  01_公用库       │   │ 02_专有库           │
              │  Public Library │   │ Private Libraries  │
              │  (authoritative)│   │  Hermes/ Codex/    │
              │  6 categories   │   │  OpenClaw/ Qoder…  │
              └──────────────┬──┘   └────────┬───────────┘
                             │               │
              ┌──────────────▼──┐   ┌────────▼───────────┐
              │ 03_交换区        │   │ 04_快照备份         │
              │ Exchange        │   │ Daily snapshots    │
              │ INBOX/OUTBOX    │   │ (versioned)        │
              └─────────────────┘   └────────────────────┘

Memory flow:

Any AI produces new memory → memory source changes
        ↓
[scheduler] coordinator.py
    scan_all.py   → capture changed memories into private libs (_scanned/)
    promote.py    → global dedup + auto-classify → write to public lib
    dispatch.py   → regenerate each AI's "shared memory injection" file
    build_index   → refresh memory index
    snapshot      → daily backup of public lib
        ↓
Each AI session start → read own 02_专有库/<AI>/共享记忆注入.md
Need details        → python scripts/search.py <keyword>
Cross-AI message    → python scripts/msg.py send --to Codex --title "..." --body "..."

Components

Script	Role
`coordinator.py`	Full pipeline orchestrator (scan→promote→dispatch→index→snapshot)
`scan_all.py`	Scan each AI's memory source (file-list mode or recursive `.md` mode)
`promote.py`	Global-hash dedup + keyword scoring classification → public library
`dispatch.py`	Generate per-AI "shared memory injection" files
`search.py`	Full-text search across public + private + exchange
`msg.py`	INBOX/OUTBOX cross-AI messaging (send / list / read)
`common.py`	Shared utilities (UTF-8/UTF-16 tolerant reading, hashing, logging)

Public Library Categories

Category	Contents
`00_用户画像`	User profile (who is the user, preferences, working style)
`01_项目知识`	Project knowledge (active projects, status, collaborators)
`02_领域知识`	Domain knowledge (research fields, technical domains)
`03_技能工具`	Skills & tools (libraries, MCP servers, tool registries)
`04_经验教训`	Lessons learned (golden rules, pitfalls, fixes)
`05_决策记录`	Decision records (ADR-style, why decisions were made)
`06_记忆索引`	Auto-generated index

Adding a New AI

Add an entry in CONFIG.json under ais (memory source path + files)
Run python scripts/coordinator.py --full
Done — the new AI gets its own private lib + injection file, and starts sharing memory

Scheduling (recommended)

# incremental: every 2 hours
python scripts/coordinator.py

# full sync + snapshot: daily at 03:00
python scripts/coordinator.py --full

Integrate with any cron scheduler (Windows Task Scheduler, systemd, Hermes cron, GitHub Actions).

Use Cases

Multi-agent households: Hermes + Codex + OpenClaw + Qoder + WorkBuddy + Claude Code all share one memory warehouse
Personal knowledge base: user profile, project knowledge, and lessons learned survive across sessions and tools
Team knowledge: one authoritative public library, every AI reads the same injected memory at session start
Automated memory ops: scheduled scan/promote/dispatch keeps the warehouse clean without manual effort

License

MIT — free for personal and commercial use. See LICENSE.

🇨🇳 中文版介绍

🎯 这是什么？

同时运行多个 AI Agent（Hermes、Codex、OpenClaw、Qoder、WorkBuddy、Claude Code…）时，每个 AI 都只拥有自己的私有记忆，结果是：

❌ 知识孤岛：一个 AI 学到的东西，其他 AI 完全看不到
❌ 重复劳动：每个 AI 都要重新学习相同的用户偏好和项目事实
❌ 无法协作：AI 之间没有办法传递消息、任务或经验
❌ 没有进化：记忆永远不会被去重、分类或版本化

AUM 把散落在各 AI 的记忆整合成一套共享、持续进化、自动维护的记忆系统：

✅ 公用库 —— 权威共享记忆（用户画像、项目知识、领域知识、经验教训、决策记录），自动分类与去重
✅ 专有库 —— 每个 AI 的原始记忆镜像 + 自动生成的注入文件（AI 每次会话启动时读取）
✅ 交换区 —— AI 之间的 INBOX/OUTBOX 消息（任务派发、通知）
✅ 调度器 —— 一条命令跑完整流程：扫描 → 提升 → 分发 → 索引 → 快照
✅ 零第三方依赖 —— 纯 Python 标准库，随处可跑

💻 图形化客户端（傻瓜式，拿来就用）

无需编程，下载安装包双击即用：

平台	下载
🪟 Windows	`AI-Unified-Memory-Windows.exe` (下载即运行)
🍎 macOS	`AI-Unified-Memory-macOS` (App)
🐧 Linux	`AI-Unified-Memory-Linux.AppImage`

界面功能：

🔄 一键同步：扫描→提升→分发→索引→快照全自动
🔍 搜索记忆：跨全库关键词检索
📊 记忆库状态：各库文件统计一目了然
⚡ 深色专业主题，全部操作一键完成

复杂的同步/分类/分发逻辑全部在后台自动运行，用户只需点一个按钮。

开发者模式

python gui_app.py                # 启动图形界面
python scripts/coordinator.py --full   # 或命令行全量同步

✨ 核心特性

🔄 一键同步：扫描 → 提升 → 分发 → 索引 → 快照，全自动流水线
🧠 记忆引擎 v2.0：
- 🔍 语义检索：关键词 + 同义词扩展 + TF 加权评分（scripts/memory_engine.py）
- 🧹 冲突消解：同主题记忆自动合并，旧版本归档 _archived/
- 🔥 热记忆：按检索频率排序，热记忆优先注入
- 📅 时间线回溯：每条记忆的版本历史
🔍 跨库搜索：公用库 + 专有库 + 交换区全文检索
📊 记忆统计：各库文件实时统计
🖥️ 图形化客户端：无需编程，下载安装包双击即用
🧠 自动维护：全局去重、关键词评分自动分类、每日快照备份
💬 AI 间消息：通过 INBOX/OUTBOX 在 AI 之间发送/读取消息
📦 零依赖：纯 Python 标准库，无需 pip 安装任何包

🚀 快速开始

# 1. 复制配置模板，填写各 AI 记忆源路径
cp CONFIG.example.json CONFIG.json

# 2. 运行全量同步循环（扫描 → 提升 → 分发 → 索引 → 快照）
python scripts/coordinator.py --full

# 3. 按需搜索共享记忆
python scripts/search.py "用户画像" --limit 10

# 4. AI 间消息
python scripts/msg.py send --to Codex --title "请审阅" --body "..."
python scripts/msg.py list
python scripts/msg.py read <消息ID>

🏗 架构

                     ┌───────────────────────────────────┐
                     │         coordinator.py            │
                     │  scan → promote → dispatch → snap │
                     └───────┬───────────────┬───────────┘
                             │               │
              ┌──────────────▼──┐   ┌────────▼───────────┐
              │  01_公用库       │   │ 02_专有库           │
              │  Public Library │   │ Private Libraries  │
              │  (authoritative)│   │  Hermes/ Codex/    │
              │  6 categories   │   │  OpenClaw/ Qoder…  │
              └──────────────┬──┘   └────────┬───────────┘
                             │               │
              ┌──────────────▼──┐   ┌────────▼───────────┐
              │ 03_交换区        │   │ 04_快照备份         │
              │ Exchange        │   │ Daily snapshots    │
              │ INBOX/OUTBOX    │   │ (versioned)        │
              └─────────────────┘   └────────────────────┘

记忆流转流程：

任意 AI 产生新记忆 → 记忆源发生变化
        ↓
[调度器] coordinator.py
    scan_all.py   → 将变更记忆采集进各专有库 (_scanned/)
    promote.py    → 全局去重 + 自动分类 → 写入公用库
    dispatch.py   → 重新生成各 AI 的「共享记忆注入」文件
    build_index   → 刷新记忆索引
    snapshot      → 每日备份公用库
        ↓
各 AI 会话启动 → 读取自己的 02_专有库/<AI>/共享记忆注入.md
需要详情        → python scripts/search.py <关键词>
AI 间消息       → python scripts/msg.py send --to Codex --title "..." --body "..."

🔧 组件说明

脚本	职责
`coordinator.py`	全流程编排器（扫描→提升→分发→索引→快照）
`scan_all.py`	扫描各 AI 记忆源（文件列表模式或递归 `.md` 模式）
`promote.py`	全局哈希去重 + 关键词评分分类 → 写入公用库
`dispatch.py`	生成各 AI 的「共享记忆注入」文件
`search.py`	公用库 + 专有库 + 交换区全文本搜索
`msg.py`	INBOX/OUTBOX 跨 AI 消息（发送 / 列表 / 读取）
`common.py`	共享工具（UTF-8/UTF-16 容错读取、哈希、日志）

📂 公用库分类

分类	内容
`00_用户画像`	用户是谁、偏好、工作方式
`01_项目知识`	进行中的项目、状态、协作者
`02_领域知识`	研究领域、技术方向
`03_技能工具`	代码库、MCP 服务器、工具注册表
`04_经验教训`	金规、踩坑记录、修复方案
`05_决策记录`	ADR 风格决策记录（为什么这么做）
`06_记忆索引`	自动生成的索引

🤝 接入新的 AI

在 CONFIG.json 的 ais 下添加条目（记忆源路径 + 文件）
运行 python scripts/coordinator.py --full
完成 —— 新 AI 自动获得专属专有库 + 注入文件，开始共享记忆

⏰ 定时调度（推荐）

# 增量同步：每 2 小时
python scripts/coordinator.py

# 全量同步 + 快照：每天 03:00
python scripts/coordinator.py --full

可接入任意定时器（Windows 任务计划程序、systemd、Hermes cron、GitHub Actions）。

🎯 使用场景

多 Agent 家庭：Hermes + Codex + OpenClaw + Qoder + WorkBuddy + Claude Code 共享同一个记忆仓库
个人知识库：用户画像、项目知识、经验教训跨会话、跨工具永久留存
团队知识沉淀：一个权威公用库，每个 AI 会话启动时读取同一份注入记忆
记忆自动化运维：定时扫描/提升/分发，无需人工维护仓库整洁

📄 许可证

MIT —— 个人与商业使用均免费。详见 LICENSE。

The memory of many AIs, unified. Each AI learns once — every AI benefits. 众多 AI 的记忆，归于一体。每个 AI 只学一次——所有 AI 共同受益。