12k+ GitHub Stars · Apache 2.0 · Python 3.13+ 12k+ GitHub Star · Apache 2.0 · Python 3.13+

The AI assistant
that never forgets. 永不遗忘的
AI 助手。

memU is a 24/7 proactive memory framework for LLMs and AI agents. It continuously captures intent, organises knowledge into a structured file hierarchy, and acts before you even ask — making always-on, evolving agents practical in production.

memU 是面向 LLM 和 AI 智能体的 24/7 主动记忆框架。它持续捕获意图，将知识组织成结构化文件层次，并在你提问之前主动行动——使始终在线、持续进化的智能体在生产环境中切实可行。

→ Get Started立即开始 ★ GitHub 20+ Bot Ideas20+ Bot 创意
92.09%
Locomo Accuracy
Locomo 准确率
~90%
Token Reduction
Token 节省
4-Layer
Memory Architecture
记忆架构
12k+ ★
GitHub Stars
GitHub Stars
Apache 2.0
Open Source Forever
永久开源
scroll to explore向下浏览
◆ Overview概览

What is memU?

什么是 memU？

memU is an open-source agentic memory framework built for 24/7 proactive agents. Released on PyPI as memu-py, it provides the persistent memory layer that LLM agents need to operate continuously without losing context — at a fraction of the cost of brute-force context windows.

memU 是专为 24/7 主动式智能体构建的开源智能记忆框架。以 memu-py 发布于 PyPI，为 LLM 智能体提供持续运行所需的持久记忆层，成本仅为暴力上下文窗口方案的零头。

🤖
24/7 Proactive — Not Reactive
24/7 主动，而非被动
Works continuously in the background without being asked. Captures intent, learns patterns, and acts proactively — even before you send a message. The agent tells you what you're about to need.
无需指令，在后台持续工作。捕获意图，学习模式，主动行动——甚至在你发消息之前。智能体会告诉你即将需要什么。
🗃️
Memory as a File System
记忆即文件系统
Structured, hierarchical, human-readable Markdown files. Browse memories like directories. Mount conversations, docs, images as queryable knowledge. Cross-link everything into a connected graph.
结构化、层次化、人类可读的 Markdown 文件。像浏览目录一样遍历记忆。将对话、文档、图像挂载为可查询知识，并交叉链接成连接图谱。
⚡
Dual Retrieval — Speed + Depth
双模检索——速度 + 深度
RAG-based embedding search for sub-100ms context assembly. LLM-based deep reading of category files for semantic understanding that embeddings can't reach. Use both together.
基于 RAG 的嵌入搜索实现亚 100ms 上下文组装。基于 LLM 的分类文件深度阅读，实现嵌入无法达到的语义理解。两者可同时使用。
💰
~90% Token Cost Reduction
约 90% Token 成本节省
Intelligent dual-mode retrieval: cheap embedding monitoring runs continuously, expensive LLM calls trigger only when needed. Running an always-on agent goes from thousands/month to tens.
智能双模检索：廉价嵌入监控持续运行，昂贵的 LLM 调用仅在必要时触发。始终在线的智能体成本从每月数千美元降至数十美元。
🔮
Intention Layer — Coming in 2026
意图层——2026 年推出
The fourth layer in development. Analyzes accumulated memories to predict what the user needs before they ask. Not just reactive assistance — cognitive anticipation. The next evolution of proactive agents.
正在开发的第四层。分析积累的记忆，在用户提问前预测其需求。不仅是被动协助——而是认知预判。主动智能体的下一次进化。
🌐
Multi-Tenant + Multimodal
多租户 + 多模态
Text, images, audio, and video all ingest natively. Each user/agent gets isolated memory under agent_id/user_id/. Enterprise-scale with per-user capacity limits and RBAC access control.
原生摄取文本、图像、音频和视频。每个用户/智能体在 agent_id/user_id/ 下获得隔离记忆。企业规模支持，带按用户容量限制和 RBAC 访问控制。
◆ Core Concept核心概念

Memory as a File System

记忆即文件系统

Just as a file system turns raw bytes into organised, navigable data — memU transforms raw interactions into structured, searchable, proactive intelligence. The analogy is literal: memories are stored as Markdown files you can open in any text editor.

正如文件系统将原始字节转化为有组织、可浏览的数据——memU 将原始交互转化为结构化、可搜索的主动智能。这个类比是字面意义的：记忆以 Markdown 文件存储，可在任何文本编辑器中打开。

File System文件系统 memU EquivalentmemU 对应 Purpose用途
📁 Folders文件夹🏷️ Memory Categories记忆分类Auto-organised topic clusters — preferences, relationships, knowledge, context自动组织的主题集群——偏好、关系、知识、上下文
📄 Files文件🧠 Memory Items记忆条目Extracted facts, preferences, skills — smallest meaningful memory unit提取的事实、偏好、技能——最小有意义的记忆单元
🔗 Symlinks符号链接🔄 Cross-references交叉引用Linked memories forming a connected knowledge graph for multi-hop reasoning链接的记忆构成多跳推理的连接知识图谱
📂 Mount Points挂载点📥 Resources资源Conversations, docs, images, audio — raw inputs feeding the memory pipeline对话、文档、图像、音频——输入记忆管道的原始资源
memory/ ├── preferences/ │ ├── communication_style.md # tone, format, conciseness preference │ ├── topic_interests.md # subjects the user engages with deeply │ └── work_schedule.md # availability, timezone, focus blocks ├── relationships/ │ ├── contacts/ │ │ ├── alice_chen.md # full context per known person │ │ └── bob_miller.md │ └── interaction_history.md # cross-contact patterns ├── knowledge/ │ ├── domain_expertise.md # deep subject matter knowledge │ └── learned_skills/ │ ├── python_debugging.md # agent-extracted skill patterns │ └── trading_strategies.md └── context/ ├── recent_conversations.md # rolling summary of recent sessions └── pending_tasks.md # open loops and uncommitted work
◆ Architecture架构

The 4-Layer Memory Architecture

四层记忆架构

Inspired by layered storage systems in computer science, memU transforms raw inputs into memory an agent can understand, retrieve, and evolve. The 4th layer — the Intention Layer — is currently in development and represents the leap from reactive to truly proactive intelligence.

受计算机科学分层存储系统启发，memU 将原始输入转化为智能体可理解、检索并演进的记忆。第四层——意图层——正在开发中，代表着从被动到真正主动智能的跨越。

Layer 4 — UPCOMINGIntention LayerComing 2026
The fourth layer currently in development. Analyzes accumulated memories to predict user needs before they are expressed. Instead of waiting for a command, the agent proactively offers suggestions, prepares answers, and initiates actions based on behavioral patterns. NevaMind-AI calls this "cognitive anticipation" — the transition from reactive AI assistants to genuinely proactive digital coworkers.
正在开发的第四层。分析积累的记忆，在用户表达需求之前预测其需要。智能体不再等待指令，而是根据行为模式主动提供建议、准备答案并发起行动。NevaMind-AI 将此称为"认知预判"——从被动 AI 助手到真正主动数字同事的转变。
▼   feeds intention up   ▼
Layer 3 — OUTPUTMemory Category Layer (Memory Files)
The aggregation layer. Memory Items are organised into structured Markdown files by category. These are the only files injected into the agent's context window during reasoning. The LLM reads them directly — no vector lookup, no chunking. Infrequently accessed memories fade from this layer but always remain retrievable from lower layers.
聚合层。记忆条目按分类组织为结构化 Markdown 文件。这些是推理期间唯一注入智能体上下文窗口的文件。LLM 直接读取——无向量查找，无分块。不常访问的记忆从此层淡出，但始终可从下层检索。
▼   extracted & clustered upward   ▼
Layer 2 — PROCESSINGMemory Item Layer
The extraction layer. The Memory Agent processes raw resources and extracts the smallest meaningful units — facts, preferences, events, relationships, skills. Items are automatically deduplicated and contradiction-resolved. The agent chooses ADD, UPDATE, or natural decay for every piece of information. Powers RAG-based fast retrieval (sub-100ms) via vector embeddings.
提取层。记忆智能体处理原始资源，提取最小有意义的单元——事实、偏好、事件、关系、技能。自动去重和解决矛盾。智能体为每条信息选择添加、更新或自然衰减。通过向量嵌入支持基于 RAG 的快速检索（亚 100ms）。
▼   preserved without modification   ▼
Layer 1 — STORAGEResource Layer (Raw Data)
The immutable foundation. Every original input — conversations, files, logs, images, audio, code — stored without compression or modification. Never pruned. Even if a memory fades from upper layers, the raw resource is always retrievable as a fallback. Full bidirectional traceability: every Category File and Memory Item links back to its source here.
不可变基础。每个原始输入——对话、文件、日志、图像、音频、代码——无压缩、无修改地存储。永不被修剪。即使记忆从上层淡出，原始资源始终可作为回退检索。完全双向可追溯：每个分类文件和记忆条目都链接回此处的源资源。
Input (text / image / audio / video / code / logs) │ ▼ ┌──────────────────────────────────────────────────────────────────────┐ │ RESOURCE LAYER — Full-fidelity raw data, never pruned │ │ Bidirectional traceability: everything links back here │ └──────────────────────┬───────────────────────────────────────────────┘ │ memU Memory Agent extracts meaning ▼ ┌──────────────────────────────────────────────────────────────────────┐ │ MEMORY ITEM LAYER — Discrete semantic units │ │ Auto-deduplicated · Conflict-resolved · Semantically linked │ │ ADD / UPDATE / DECAY decision per item · RAG retrieval (sub-100ms) │ └──────────────────────┬───────────────────────────────────────────────┘ │ clustered into topic categories ▼ ┌──────────────────────────────────────────────────────────────────────┐ │ MEMORY CATEGORY LAYER — Structured Markdown files by topic │ │ LLM reads directly · Injected into context · Self-evolving │ └──────────────────────┬───────────────────────────────────────────────┘ │ behavioral pattern analysis ▼ ┌──────────────────────────────────────────────────────────────────────┐ │ INTENTION LAYER [IN DEVELOPMENT] — Cognitive Anticipation │ │ Predicts needs before expressed · Proactive action triggers │ └──────────────────────────────────────────────────────────────────────┘ │ ▼ AGENT CONTEXT WINDOW → reasoning → response → action
◆ How It Works工作原理

The Proactive Memory Lifecycle

主动记忆生命周期

memU runs a continuous sync loop between the main agent and the memU memory bot. While the agent handles queries, memU observes every interaction, extracts insights, predicts next steps, and pre-loads context — all in parallel, all without being asked.

memU 在主智能体和 memU 记忆机器人之间运行持续同步循环。主智能体处理查询时，memU 并行观察每次交互、提取洞察、预测下一步并预加载上下文——全程无需指令。

USER QUERY / TRIGGER │ ┌────────────────────┴────────────────────┐ ▼ ▼ ┌──────────────────────────┐ ┌──────────────────────────────────┐ │ 🤖 MAIN AGENT │ │ 🧠 MEMU BOT │ │ Handle queries + tasks │ ◄──► │ Monitor, memorize, anticipate │ ├──────────────────────────┤ ├──────────────────────────────────┤ │ 1. Parse query + intent │ ───► │ 1. Observe interactions │ │ │ │ Track conversation flow │ │ 2. Plan + Execute │ ◄─── │ 2. Extract + store memories │ │ Call tools, generate │inject│ Dedup, resolve conflicts │ │ │memory│ Update user/agent profile │ │ 3. Respond to user │ ───► │ 3. Predict user intent │ │ │ │ Anticipate next needs │ │ 4. Loop / await │ ◄─── │ 4. Run proactive tasks │ │ or proactive hint │suggest Pre-fetch relevant context │ └──────────────────────────┘ │ Prepare recommendations │ └──────────────────────────────────┘ └──────────────────────────────────────────┘ CONTINUOUS SYNC LOOP Agent ◄──► MemU Bot ◄──► PostgreSQL
◆ Retrieval检索

Dual retrieval — speed + depth

双模检索——速度 + 深度

Two complementary retrieval strategies that can be used independently or combined. RAG for always-on monitoring, LLM for deep anticipatory reasoning.

两种互补的检索策略，可独立使用或组合。RAG 用于始终在线监控，LLM 用于深度预判推理。

Mode 1RAG RetrievalRAG 检索
Operates at the Memory Item Layer. Fast vector embedding similarity search — zero LLM calls. Ideal for continuous background monitoring and high-frequency queries. The cheap monitoring backbone that runs 24/7.
在记忆条目层运行。快速向量嵌入相似度搜索——零 LLM 调用。适合持续后台监控和高频查询。24/7 运行的廉价监控骨干。
⚡ Sub-100msEmbedding only
Mode 2LLM Deep RetrievalLLM 深度检索
LLM directly reads entire Memory Category Files. Automatically refines search queries, performs multi-hop reasoning across the memory graph, and terminates early when sufficient context is gathered. Catches what embeddings miss.
LLM 直接读取完整的记忆分类文件。自动优化搜索查询，在记忆图谱上执行多跳推理，收集到足够上下文后提前终止。捕获嵌入遗漏的内容。
🧠 Deep semanticComplex queries
python — retrieve()
# RAG — fast, continuous monitoring result = await service.retrieve( queries=[{"role": "user", "content": {"text": "What are Alice's preferences?"}}], where={"user_id": "alice"}, method="rag" # sub-100ms, zero LLM cost ) # LLM — deep reasoning for complex multi-hop queries result = await service.retrieve( queries=[{"role": "user", "content": {"text": "What career decisions has she made, and how do they relate to her team dynamics?"}}], where={"user_id": "alice"}, method="llm" # reads category files, multi-hop reasoning )
◆ Multimodal多模态

Memory for any input type

支持任意输入的记忆

memU ingests text, images, audio, and video natively — all preserved at full fidelity in the Resource Layer and progressively unified into text memory items for consistent LLM reasoning.

memU 原生摄取文本、图像、音频和视频——全部以完整保真度保存在资源层，并逐步统一为文本记忆条目，以实现一致的 LLM 推理。

python — memorize() modalities
result = await service.memorize( resource_url="meeting_2026_q1.mp3", modality="audio", # "conversation" | "document" | "image" | "video" | "audio" user={"user_id": "alice"} ) # Audio → transcribed → facts + preferences extracted + categories updated # Image → vision model describes → visual context linked to conversations # Video → key frames + audio → unified text memory # Code → patterns, decisions, errors extracted as memory items # All modalities → Resource Layer (full fidelity) → Item Layer → Category Layer
◆ Benchmarks基准测试

92.09% on Locomo — state of the art

Locomo 基准 92.09%——业界领先

memU achieves 92.09% average accuracy on the Locomo benchmark across all reasoning tasks — demonstrating reliable production-grade memory operations including fact recall, temporal reasoning, preference tracking, and relationship inference.

memU 在 Locomo 基准的全部推理任务中实现 92.09% 平均准确率——在事实回忆、时序推理、偏好追踪和关系推理等生产级记忆操作中表现可靠。

🏆
What Locomo measures: Long-term conversation memory quality — the ability to recall specific facts, preferences, and relationships from hundreds of prior conversations, including temporal ordering and entity relationship inference. 92.09% accuracy is best-in-class for self-hosted, local-first memory frameworks.
Locomo 衡量什么：长期对话记忆质量——从数百次先前对话中回忆具体事实、偏好和关系的能力，包括时序排序和实体关系推理。92.09% 准确率在自托管、本地优先的记忆框架中处于最高水平。
→ View full experiment data查看完整实验数据
◆ Getting Started开始使用

Cloud or Self-Hosted

云端或自托管

Two paths. Cloud (memu.so) for instant start with zero infrastructure. Self-hosted for full data sovereignty, offline operation, and unlimited customisation — free forever under Apache 2.0.

两条路径。云端（memu.so）可零基础设施立即启动。自托管提供完整数据主权、离线运行和无限定制——Apache 2.0 下永久免费。

☁️
Cloud (memu.so)
云端（memu.so）
Hosted service with 7×24 continuous learning. No infrastructure to manage. REST API v3 with usage-based pricing and a free starting tier. Enterprise plans with custom proactive workflows available.
提供 7×24 持续学习的托管服务。无需管理基础设施。REST API v3 按用量计费，提供免费起步层。支持自定义主动工作流的企业计划。
→ Try Free免费体验
🖥️
Self-Hosted (Open Source)
自托管（开源）
Full data sovereignty. Works 100% offline with Ollama. Supports in-memory storage for dev, PostgreSQL + pgvector for production, and memU-server + Temporal for enterprise workflow deployments.
完整数据主权。配合 Ollama 100% 离线运行。开发支持内存存储，生产支持 PostgreSQL + pgvector，企业工作流部署支持 memU-server + Temporal。
01

Install memU

安装 memU

Requires Python 3.13+. Install via PyPI. For development, clone and run make install which sets up venv, all dependencies, and pre-commit hooks (Ruff, Black, mypy, deptry).

需要 Python 3.13+。通过 PyPI 安装。开发模式下克隆仓库并运行 make install，自动配置 venv、所有依赖和预提交钩子（Ruff、Black、mypy、deptry）。

terminal
pip install memu-py # Or for development (with test suite and pre-commit hooks) git clone https://github.com/NevaMind-AI/memU.git cd memU && make install
02

Choose storage backend

选择存储后端

In-memory for development — zero setup, data lives in RAM. PostgreSQL + pgvector for production — persistent, scalable, supports multi-agent deployments.

开发用内存存储——零配置，数据存在 RAM 中。生产用 PostgreSQL + pgvector——持久化、可扩展，支持多智能体部署。

terminal — zero setup
export OPENAI_API_KEY=your_key cd tests && python test_inmemory.py
terminal — PostgreSQL + pgvector
docker run -d \ --name memu-postgres \ -e POSTGRES_USER=postgres \ -e POSTGRES_PASSWORD=postgres \ -e POSTGRES_DB=memu \ -p 5432:5432 \ pgvector/pgvector:pg16 export OPENAI_API_KEY=your_key cd tests && python test_postgres.py
03

Deploy memU-server + Temporal (enterprise)

部署 memU-server + Temporal（企业级）

For teams that need a REST API layer, background job orchestration, and a management UI — deploy memU-server with the full Docker Compose stack. This starts PostgreSQL, the Temporal workflow engine (gRPC on port 7233), and the Temporal UI (port 8088).

需要 REST API 层、后台作业编排和管理界面的团队，可使用完整 Docker Compose 栈部署 memU-server。这将启动 PostgreSQL、Temporal 工作流引擎（gRPC 端口 7233）和 Temporal UI（端口 8088）。

terminal — full enterprise stack
git clone https://github.com/NevaMind-AI/memU-server.git cd memU-server export OPENAI_API_KEY=your_key # Quick start (FastAPI only) uv sync && uv run fastapi dev # Full stack: PostgreSQL + Temporal + UI docker compose up -d # Services: # PostgreSQL → port 5432 (memory store) # Temporal → port 7233 (workflow engine gRPC) # Temporal UI → port 8088 (web management interface)
04

Configure LLM provider(s)

配置 LLM 提供商

memU supports named LLM profiles — use different models for different operations. Power tip: assign a fast cheap model for extraction/monitoring and an expensive model only for deep reasoning. This alone cuts monthly costs 60–70%.

memU 支持命名 LLM 配置——为不同操作使用不同模型。进阶技巧：为提取/监控分配廉价快速模型，仅在深度推理时使用昂贵模型。仅此一项即可降低 60–70% 的月度成本。

OpenAI GPT-4o / o3 Anthropic Claude Opus/Sonnet Google Gemini 2.5 DeepSeek V3 / R1 Qwen3 / Grok OpenRouter (any model) Ollama — Free & Offline Voyage AI (embeddings)
python — multi-profile LLM config
from memu import MemUService service = MemUService( llm_profiles={ # Fast cheap model: extraction, monitoring, routine queries "default": { "base_url": "https://api.openai.com/v1", "api_key": "sk-...", "chat_model": "gpt-4o-mini", # cheap for monitoring "client_backend": "sdk" }, # Premium model: deep reasoning and complex anticipation only "reasoning": { "base_url": "https://api.anthropic.com", "api_key": "sk-ant-...", "chat_model": "claude-opus-4-5" # premium for complex tasks }, # Separate embedding provider for maximum retrieval quality "embedding": { "base_url": "https://api.voyageai.com/v1", "api_key": "pa-...", "embed_model": "voyage-3.5-lite" } } )
05

Run your first proactive workflow

运行第一个主动工作流

Test the full memorize → retrieve loop, then graduate to the proactive example which demonstrates intent prediction and autonomous action.

测试完整的记忆 → 检索循环，然后尝试主动示例，演示意图预测和自主行动。

terminal
# Test full memory loop python examples/example_1_conversation_memory.py # Test skill extraction (self-improving agent) python examples/example_2_skill_extraction.py # Test multimodal memory python examples/example_3_multimodal_memory.py # Test proactive mode — agent acts WITHOUT being asked cd examples/proactive && python proactive.py
◆ Provider Integration提供商集成

OpenRouter + Multi-Profile Setup

OpenRouter + 多配置设置

OpenRouter gives access to hundreds of LLM models through a single API. Combined with memU's multi-profile system, you can run the cheapest model for monitoring and the most capable for deep reasoning — with complete cost control.

OpenRouter 通过单一 API 提供对数百个 LLM 模型的访问。结合 memU 的多配置系统，你可以用最廉价的模型进行监控，用最强大的模型进行深度推理——完全掌控成本。

python — OpenRouter config
from memu import MemoryService service = MemoryService( llm_profiles={ "default": { "provider": "openrouter", "client_backend": "httpx", "base_url": "https://openrouter.ai", "api_key": "sk-or-...", "chat_model": "anthropic/claude-sonnet-4-5", "embed_model": "openai/text-embedding-3-small", }, }, database_config={"metadata_store": {"provider": "inmemory"}}, )
◆ Platforms平台

Works where your team already is

在团队已有平台运行

✈️
Telegram
✓ Live
Full Bot API, groups, channels
完整 Bot API，群组，频道
🎮
Discord
✓ Live
Slash commands, threads
斜杠命令，线程
💬
Slack
✓ Live
Workspace apps, DMs
工作区应用，私信
🪶
Feishu / Lark
✓ Live
Native bots, group chats
原生机器人，群聊
📧
Email
Soon
📱
WhatsApp
Soon
🌐
Web UI
Soon
⌨️
CLI
✓ Live
Developer terminal interface
开发者终端界面
◆ API ReferenceAPI 参考

Core APIs — memorize() + retrieve()

核心 API — memorize() + retrieve()

Two primary operations. memorize() processes any input and immediately updates all three memory layers — zero delay. retrieve() queries with RAG or LLM mode and returns categorised results with full source traceability.

两个主要操作。memorize() 处理任意输入并立即更新全部三层记忆——零延迟。retrieve() 以 RAG 或 LLM 模式查询，返回带完整来源追溯的分类结果。

python — memorize()
result = await service.memorize( resource_url="conversations/session_042.json", modality="conversation", user={"user_id": "alice"} ) # Immediate response — zero delay: result["resource"] # stored metadata result["items"] # extracted memory items (ready now) result["categories"] # updated category structure # ✓ Auto-categorisation ✓ Deduplication ✓ Conflict resolution
python — retrieve() with proactive filtering
# User-scoped (most common) result = await service.retrieve( queries=[{"role": "user", "content": {"text": "What are her preferences?"}}], where={"user_id": "alice"}, method="rag" ) # Multi-agent coordination result = await service.retrieve( queries=[{"role": "user", "content": {"text": "What does the team know?"}}], where={"agent_id__in": ["agent_1", "agent_2"]}, method="llm" ) # Response structure: result["categories"] # relevant topic areas, auto-prioritised result["items"] # specific memory facts result["resources"] # original sources for traceability result["next_step_query"] # predicted follow-up context
◆ Cloud API (v3)
MethodEndpointDescription描述
POST/api/v3/memory/memorizeRegister continuous learning task注册持续学习任务
GET/api/v3/memory/memorize/status/{id}Check real-time processing status查看实时处理状态
POST/api/v3/memory/categoriesList auto-generated categories列出自动生成的分类
POST/api/v3/memory/retrieveQuery memory with proactive context loading查询记忆（支持主动上下文加载）
cloud API base
Base URL: https://api.memu.so Auth: Authorization: Bearer YOUR_API_KEY
◆ Framework Integrations框架集成
Framework框架Integration集成方式Status
LangChainDrop-in for ConversationBufferMemory替换 ConversationBufferMemory✓
LangGraphState-based memory node in graph workflows图工作流中的状态记忆节点✓
CrewAIMemUMemoryStore replaces default memoryMemUMemoryStore 替换默认记忆✓
OpenAI Agents SDKCross-session persistence via endpoints通过端点实现跨会话持久化✓
Claude Agent SDKPersistence alongside Anthropic's tool loopAnthropic 工具循环旁的持久化层✓
n8nWorkflow automation with persistent memory带持久记忆的工作流自动化✓
DifyMemory backend for Dify agentsDify 智能体记忆后端✓
LlamaIndexPersistent agent memory layer持久智能体记忆层✓
AutoGPTLong-term memory for autonomous agents自主智能体的长期记忆✓
OpenRouterAny model via unified API通过统一 API 使用任意模型✓
AutoGenShared memory across multi-agent conversations跨多智能体对话的共享记忆In Progress
REST API / Go SDKLanguage-agnostic HTTP + Go client语言无关 HTTP + Go 客户端✓
◆ Comparisons对比

memU vs mem0

memU vs mem0

The most common question. Core difference: mem0 is a RAG system. memU is a true memory layer. They solve fundamentally different problems — using the wrong one for your use case leads to both poor quality and unnecessary cost.

最常见的问题。核心区别：mem0 是 RAG 系统，memU 是真正的记忆层。它们解决根本不同的问题——为错误的用例选择错误的工具会导致质量差和不必要的成本。

⚠
Use the right tool: memU for agent memory (preferences, history, behavioral patterns). RAG (LangChain, LlamaIndex) for large document search. They complement each other — many production systems use both in the same agent pipeline.
使用正确的工具：memU 用于智能体记忆（偏好、历史、行为模式）。RAG（LangChain、LlamaIndex）用于大规模文档搜索。两者相互补充——许多生产系统在同一智能体流水线中同时使用两者。
Dimension维度memUmem0mem0
PARADIGM范式True memory layer — agentic extraction, 4-layer file hierarchy, self-evolving structure真正的记忆层——智能体提取，四层文件层次，自进化结构RAG system — chunks, embeds, similarity searchRAG 系统——分块、嵌入、相似度搜索
STORAGE存储✓ Clean structured Markdown — no chunking noise, zero irrelevant fragments✓ 干净的结构化 Markdown——无分块噪声，零无关碎片Chunked docs in vector store — retrieval noise向量存储中的分块文档——检索噪声
RETRIEVAL检索✓ Dual-mode: sub-100ms RAG + LLM direct file reading✓ 双模：亚 100ms RAG + LLM 直接文件阅读Embedding similarity only — misses relationships仅嵌入相似度——遗漏关系
DEDUP去重✓ Automatic on every write — ADD/UPDATE/DECAY per item✓ 每次写入自动——每条目 ADD/UPDATE/DECAYManual or none手动或无
PROACTIVE主动性✓ 24/7 — Intention Layer predicts needs before expressed✓ 24/7——意图层在需求表达前预测Reactive only仅被动响应
BENCHMARK基准✓ 92.09% Locomo accuracy — best-in-class local-first✓ Locomo 92.09%——本地优先最高水准Different benchmark focus (RAG pipelines)不同基准侧重（RAG 流程）
TRANSPARENCY透明度✓ Plain Markdown files — open any memory in a text editor✓ 普通 Markdown 文件——在文本编辑器中打开任意记忆Opaque vectors不透明向量
MIGRATION迁移✓ memU provides migration tools to import existing mem0 memories✓ memU 提供工具导入现有 mem0 记忆——
◆ Comparison对比

memU vs OpenClaw memory

memU vs OpenClaw 记忆

OpenClaw has 200k+ GitHub stars. We respect it deeply. memU is not competing with OpenClaw's orchestration layer — it replaces OpenClaw's native memory backend and makes every OpenClaw agent dramatically smarter over time. Several OpenClaw users have reported significantly improved recall after switching to memU, particularly for cross-project queries and long-term preference tracking.

OpenClaw 有 20 万+ GitHub Star。我们深深尊重它。memU 不与 OpenClaw 的编排层竞争——而是替换其原生记忆后端，让每个 OpenClaw 智能体随时间显著变得更智能。多位 OpenClaw 用户报告在切换到 memU 后记忆召回显著改善，尤其在跨项目查询和长期偏好追踪方面。

DimensionmemUOpenClaw nativeOpenClaw 原生
STORAGE存储4-layer structured graph, Markdown files by category四层结构化图谱，按分类的 Markdown 文件Flat files + SQLite vector store平面文件 + SQLite 向量存储
OVERFLOW溢出✓ Auto-flush — raw data always preserved✓ 自动刷新——原始数据永久保留Compaction summarisation — detail loss压缩摘要——细节丢失
PROACTIVE主动性✓ 24/7 intent capture + Intention Layer✓ 24/7 意图捕获 + 意图层Reactive to commands only仅响应命令
TOKEN COSTToken 成本✓ ~90% reduction via targeted injection✓ 精准注入减少约 90%Full history grows linearly完整历史线性增长
BENCHMARK基准✓ 92.09% Locomo✓ Locomo 92.09%No published memory benchmark无已发布记忆基准
◆ Comparison对比

memU vs RAG vs Context Window

memU vs RAG vs 上下文窗口

ApproachmemURAGRAGFull Context完整上下文
BEST FOR最适合Agent memory — preferences, learned patterns, long-term context智能体记忆——偏好、学习模式、长期上下文Searching large static document corpora搜索大型静态文档语料库Short single-session tasks短期单会话任务
SCALE扩展性Unlimited — only relevant subset injected无限制——仅注入相关子集Scales with vector DB随向量数据库扩展Hard token limit硬 Token 限制
COST成本Low — dual-mode: cheap RAG + LLM only when needed低——双模：廉价 RAG + 按需 LLMMedium — embedding + tokens per query中——每次查询嵌入 + TokenHigh — full history every call高——每次调用完整历史
PROACTIVE主动性Yes — Intention Layer predicts needs before asked是——意图层在提问前预测需求Reactive only仅被动Reactive only仅被动
DETAIL LOSS细节丢失None — raw resources never deleted无——原始资源永不删除Chunking loses cross-fragment context分块丢失跨片段上下文Compaction causes permanent loss压缩导致永久丢失
◆ Real-World Scenarios真实场景

memU in action — proactive scenarios

memU 实战——主动场景

These are real examples from the official memU repository. Each demonstrates memU running in the background, observing, learning, and acting without being explicitly asked.

来自 memU 官方仓库的真实示例。每个示例展示 memU 在后台运行，在没有明确指令的情况下观察、学习和行动。

MemU tracks reading history, saved articles, and search queries continuously. When new content arrives, the agent surfaces it proactively — including papers by authors the user has previously cited — without any prompt.

memU 持续追踪阅读历史、收藏文章和搜索查询。当新内容出现时，智能体主动推送——包括用户之前引用过的作者发表的论文——无需任何提示。

example output
Agent: "I found 3 new papers on RAG optimisation matching your research. Dr. Chen (whom you've cited before) published yesterday. Want me to summarise them and add to your knowledge base?"

Agent learns email response patterns, priority contacts, scheduling preferences, and writing style. Proactively drafts replies, categorises inbox, and detects scheduling conflicts across threads — all without being asked.

智能体学习邮件回复模式、优先联系人、日程偏好和写作风格。主动起草回复、整理收件箱，并检测跨线程的日程冲突——全部无需指令。

example output
Agent: "12 new emails. Drafted responses for 3 routine requests. Flagged 2 urgent items from priority contacts. John mentioned a conflict — should I reschedule tomorrow's meeting?"

Learns risk tolerance, preferred sectors, and behavioral patterns from past trades. Alerts are contextualised with portfolio state and personal history — not just raw price thresholds.

从过往交易学习风险承受度、偏好板块和行为模式。警报结合投资组合状态和个人历史进行上下文化——不仅仅是原始价格阈值。

example output
Agent: "NVDA dropped 5% after hours. Based on your history, you typically buy tech dips above 3%. Current allocation allows $2,000 more while maintaining your 70/30 target."

Agent monitors its own execution logs and outcomes. Identifies success/failure patterns. Auto-generates skill guides from experience. Proactively suggests better strategies for similar future tasks — the agent genuinely gets smarter over time.

智能体监控自身执行日志和结果。识别成功/失败模式。从经验自动生成技能指南。为类似未来任务主动建议更好的策略——智能体随时间真正变得更智能。

terminal
python examples/example_2_skill_extraction.py # ✓ Monitors actions + outcomes continuously # ✓ Identifies patterns in successes and failures # ✓ Auto-generates skill guides from experience # ✓ Suggests strategies for similar future tasks
◆ Power User Tips进阶技巧

6 memU Power Lifehacks

6 个 memU 进阶技巧

Practical tips from production deployments. These optimisations can cut costs by 60–70% and dramatically improve retrieval accuracy without changing your core architecture.

来自生产部署的实用技巧。这些优化可在不改变核心架构的情况下将成本降低 60–70%，并显著提升检索准确性。

01
Seed Memory Strategically — Don't Wait for Organic Growth
策略性播种记忆——不要等待自然积累
Before deploying to production, manually seed critical knowledge: import existing documents, customer profiles, technical runbooks, and historical incident data. This gives your agent a running start instead of a cold start. A well-seeded agent reaches useful behavior in hours rather than weeks.
在生产部署前，手动播种关键知识：导入现有文档、客户资料、技术运行手册和历史事故数据。这让你的智能体从跑步开始而非从冷启动。良好播种的智能体在数小时内而非数周内达到有效行为。
02
Use where Filters for Targeted Proactive Monitoring
使用 where 过滤器进行精准主动监控
The where clause in retrieve() scopes continuous monitoring to specific data streams. Instead of watching everything (expensive), use targeted filters to focus on the signals that matter most. A well-tuned where filter reduces monitoring costs by 40–60% while keeping proactive behavior sharp for high-value events.
retrieve() 中的 where 子句将持续监控范围限定到特定数据流。不要监控所有内容（成本高），而是用精准过滤器聚焦最重要的信号。精心调优的 where 过滤器在降低 40–60% 监控成本的同时，保持对高价值事件的主动响应锐度。
03
Assign Cheap Models for Monitoring, Expensive for Reasoning
监控用廉价模型，推理用昂贵模型
Use memU's named LLM profiles to assign GPT-4o-mini or Claude Haiku for routine extraction and monitoring (runs 24/7), and reserve Claude Opus or GPT-4 for deep reasoning triggers only. Track your monitoring-to-reasoning call ratio — if LLM reasoning triggers too frequently, refine your memory categories. This optimization alone cuts monthly costs 60–70%.
使用 memU 的命名 LLM 配置，将 GPT-4o-mini 或 Claude Haiku 分配给常规提取和监控（24/7 运行），仅将 Claude Opus 或 GPT-4 保留给深度推理触发。追踪你的监控与推理调用比率——如果 LLM 推理触发过于频繁，优化你的记忆分类。仅此优化即可降低每月成本 60–70%。
04
Version Control Your Memory Files
对记忆文件进行版本控制
Since memU stores memories as Markdown files, put them under Git version control. This gives you a complete audit trail of what your agent knows, when it learned something, and how its knowledge evolved. It also enables easy rollback if a bad memory injection corrupts agent behavior — just git revert the offending commit.
由于 memU 将记忆存储为 Markdown 文件，可以将其纳入 Git 版本控制。这提供智能体知识的完整审计追踪——知道什么、何时学习、知识如何演进。还支持在错误记忆注入破坏智能体行为时轻松回滚——只需 git revert 相关提交。
05
Schedule Memory Hygiene Routines
安排记忆清理例程
Over time, memory stores accumulate outdated or contradictory information. While memU naturally decays unused memories, actively schedule weekly cleanup routines where the agent reviews and consolidates its knowledge base. Use the built-in conflict detection to surface contradictions before they cause hallucination-style errors in retrieval.
随着时间推移，记忆库会积累过时或矛盾的信息。虽然 memU 会自然衰减未使用的记忆，但仍应主动安排每周清理例程，让智能体审查和整合其知识库。使用内置冲突检测在矛盾导致检索中出现幻觉式错误之前将其暴露出来。
06
Build memMesh for Multi-Agent Collective Intelligence
构建 memMesh 实现多智能体集体智能
The community-built memMesh architecture connects multiple specialized agents through a shared memU PostgreSQL memory store. When one agent (researcher, builder, reviewer) discovers a fact, it becomes immediately available to all others. Past failures are automatically checked against new proposals, preventing regression and enabling genuine organizational learning across your entire agent fleet.
社区构建的 memMesh 架构通过共享的 memU PostgreSQL 记忆存储连接多个专业化智能体。当一个智能体（研究者、构建者、审查者）发现一个事实时，它立即对所有其他智能体可用。过去的失败自动与新提案对照检查，防止回归，实现整个智能体团队的真正组织学习。
◆ 20+ Bot Ideas20+ Bot 创意

What to build with memU

用 memU 构建什么

memU enables a new class of agent that gets genuinely smarter over time. These aren't chatbots — they're always-on agents that accumulate institutional knowledge and act proactively. Here are 20+ concrete ideas across every industry.

memU 开创了一类随时间真正变得更智能的智能体新形态。这些不是聊天机器人——而是积累机构知识并主动行动的始终在线智能体。以下是跨行业的 20+ 具体创意。

Team Ops
团队运营
🔁 Daily Standup Bot
🔁 每日站会 Bot
Collects updates, summarises blockers, posts a digest every morning. Remembers task context across weeks — not just today.
收集更新，汇总阻碍，每天早上发布摘要。跨周记住任务上下文——不仅限于今天。
  • Detects recurring blockers and auto-escalates
  • 检测反复阻碍并自动升级
  • Weekly velocity trend with pattern analysis
  • 带模式分析的每周速度趋势
Engineering
工程研发
🚨 Incident Response Bot
🚨 事故响应 Bot
Monitors channels for production alerts. Assembles incident history, links past root causes, coordinates team, auto-generates post-mortems.
监控频道中的生产告警。整合事故历史，关联过往根本原因，协调团队，自动生成事后复盘。
  • Runbook knowledge from every past incident
  • 来自每次过往事故的运行手册知识
  • Learns which change types cause regressions
  • 学习哪类变更会导致回归
Sales / CRM
销售/CRM
🤝 CRM Intelligence Bot
🤝 CRM 智能 Bot
Remembers every customer touchpoint across email, Slack, and calls. Surfaces context before meetings, detects deal risk via activity silence.
记住跨邮件、Slack 和通话的每个客户接触点。会议前提供上下文，通过活动沉默检测交易风险。
  • Personalised pre-call briefs from full history
  • 来自完整历史的个性化通话前简报
  • Win/loss pattern coaching for reps
  • 为销售代表提供赢损模式辅导
Knowledge Mgmt
知识管理
📚 Team Knowledge Bot
📚 团队知识 Bot
Silently indexes all channel decisions, rationale, and discussions. When someone asks a question answered 6 months ago, this bot knows.
静默索引所有频道决策、理由和讨论。当有人提出 6 个月前已回答的问题时，Bot 知道答案。
  • Answers "why did we do X?" instantly
  • 即时回答"我们为何这样做？"
  • New-hire onboarding with institutional memory
  • 用机构记忆支撑新员工入职
Finance
金融
📊 Trading Intelligence Bot
📊 交易智能 Bot
Learns risk tolerance and behavioral patterns from executed trades. Alerts are personalised and contextualised — not just raw price thresholds.
从已执行交易学习风险承受度和行为模式。警报个性化且上下文化——不仅仅是原始价格阈值。
  • Learns from executed vs. ignored alerts
  • 从已执行 vs 忽略的警报中学习
  • Anticipates tax-loss harvesting windows
  • 预判税损收割时机
Personal
个人
🧠 Second Brain Bot
🧠 第二大脑 Bot
Always-on personal assistant — remembers every note, task, goal, and conversation. Learns work patterns and energy cycles.
始终在线的个人助手——记住每个笔记、任务、目标和对话。学习工作模式和精力周期。
  • Daily briefing tailored to current projects
  • 根据当前项目定制每日简报
  • Connects ideas across months of conversations
  • 连接数月对话中的想法
DevOps
DevOps
🚀 Release Intelligence Bot
🚀 发布智能 Bot
Tracks every deploy, links to PRs and issues, learns which change types historically cause regressions. Predicts risk level of upcoming releases.
跟踪每次部署，与 PR 和 issue 关联，学习哪类历史变更会导致回归。预测即将发布的风险级别。
  • Full deploy → regression history in memory
  • 记忆中完整的部署 → 回归历史
  • Auto-generates release notes from commits
  • 从提交自动生成发布说明
HR / Ops
HR / 运营
🎓 Onboarding Navigator Bot
🎓 入职引导 Bot
Guides new hires with memory of their individual progress. Builds FAQ from repeated onboarding questions automatically.
记住每位新员工个人进度的引导助手。自动从重复入职问题构建 FAQ。
  • Tracks checklist completion per person
  • 追踪每人的清单完成情况
  • Connects new hires to right internal experts
  • 将新员工与合适的内部专家对接
Research
研究
🔬 Research Companion Bot
🔬 研究伴侣 Bot
Builds a knowledge graph of your research over time. Connects new papers to existing notes, detects contradictions across sources, and surfaces relevant findings proactively.
随时间构建你研究的知识图谱。将新论文连接到现有笔记，检测来源间矛盾，主动推送相关发现。
  • Weekly "what you might have missed" digest
  • 每周"你可能错过的"摘要
  • Author relationship and citation tracking
  • 作者关系和引用追踪
Companion
陪伴
💙 AI Companion Bot
💙 AI 陪伴 Bot
Maintains emotional continuity across months of conversation. Detects behavioral changes, follows up proactively, builds deep personalized connection over time.
在数月对话中保持情感连续性。检测行为变化，主动跟进，随时间建立深度个性化连接。
  • Notices patterns: "You've been online late again"
  • 注意到模式："你又很晚上线了"
  • Follows up on absent users proactively
  • 主动跟进长时间未活跃的用户
Customer Support
客户支持
🎧 Support Memory Bot
🎧 支持记忆 Bot
Transforms isolated tickets into ongoing relationships. Remembers customer preferences, past resolutions, and communication style. Proactively prevents churn by detecting frustration patterns early.
将孤立的工单转化为持续的关系。记住客户偏好、过往解决方案和沟通风格。通过早期检测挫败模式主动防止客户流失。
  • Cross-platform memory: tickets + email + chat
  • 跨平台记忆：工单 + 邮件 + 聊天
  • Proactive discount triggers prevented a $720 MRR churn (documented)
  • 主动折扣触发防止了 720 美元 MRR 流失（有记录）
Education
教育
📖 AI Tutor Bot
📖 AI 辅导 Bot
Remembers every student's learning gaps, preferred explanation style, and prior sessions. Adapts difficulty and format dynamically. Resurfaces forgotten topics proactively before tests.
记住每位学生的知识缺口、偏好的解释风格和过往课程。动态调整难度和格式。在考试前主动重新介绍遗忘的主题。
  • Per-student knowledge gap tracking
  • 按学生追踪知识缺口
  • Spaced repetition driven by memory access patterns
  • 由记忆访问模式驱动的间隔重复
Health / Wellness
健康/健身
🏃 Fitness Coach Bot
🏃 健身教练 Bot
Tracks workout history, nutrition patterns, energy levels, and recovery. Adapts training plans based on actual performance over months, not preset templates.
追踪训练历史、营养模式、能量水平和恢复情况。基于数月的实际表现（而非预设模板）调整训练计划。
  • Detects overtraining signals before injury
  • 在受伤前检测过度训练信号
  • Correlates nutrition with performance outcomes
  • 将营养与表现结果关联
Legal
法律
⚖️ Legal Research Bot
⚖️ 法律研究 Bot
Builds a knowledge graph of case law, precedents, and client matters over time. Surfaces relevant precedents proactively when new cases arrive. Learns each attorney's research preferences.
随时间构建判例法、先例和客户事项的知识图谱。当新案件到来时主动推送相关先例。学习每位律师的研究偏好。
  • Cross-case precedent linking
  • 跨案例先例关联
  • Tracks jurisdiction-specific rule changes
  • 追踪特定司法管辖区的规则变化
Healthcare
医疗健康
🏥 Care Coordination Bot
🏥 护理协调 Bot
Maintains patient context across providers, appointments, and care events. Proactively surfaces relevant history when new providers join a case. Tracks medication adherence and appointment follow-through.
跨提供者、预约和护理事件维护患者上下文。当新提供者加入案例时主动推送相关历史。追踪用药依从性和预约跟进情况。
  • Cross-provider patient history surfacing
  • 跨提供者患者历史推送
  • Proactive medication adherence monitoring
  • 主动用药依从性监控
Security
安全
🔒 Security Threat Bot
🔒 安全威胁 Bot
Builds a persistent knowledge base of security incidents, CVEs, and system behavior baselines. Detects anomalies by comparing current activity to historical patterns — not just rules.
构建安全事件、CVE 和系统行为基线的持久知识库。通过与历史模式比较（而非仅靠规则）检测异常。
  • Baseline deviation detection over months
  • 数月维度的基线偏差检测
  • Links current threats to past incident patterns
  • 将当前威胁与过往事件模式关联
Real Estate
房地产
🏠 Property Scout Bot
🏠 房产搜索 Bot
Learns buyer preferences across dozens of property viewings — what they actually reacted to versus stated preferences. Proactively alerts on new listings that match the real learned criteria.
通过数十次看房学习买家偏好——他们的实际反应与陈述偏好的差异。主动提醒符合真实学习标准的新房源。
  • Learned vs stated preference differentiation
  • 区分学习到的偏好与陈述的偏好
  • Tracks neighborhood trend context over time
  • 随时间追踪街区趋势上下文
Recruiting
招聘
🎯 Recruiting Intelligence Bot
🎯 招聘智能 Bot
Builds deep candidate profiles across all interactions. Remembers every interview observation, feedback thread, and hiring decision. Proactively resurfaces past candidates when new roles open.
在所有交互中构建深度候选人档案。记住每次面试观察、反馈线程和招聘决定。当新岗位开放时主动重新推送过往候选人。
  • Pattern-matches candidates to open roles
  • 将候选人与开放岗位进行模式匹配
  • Tracks hiring decision quality over time
  • 随时间追踪招聘决策质量
IoT / Hardware
物联网/硬件
🏭 Predictive Maintenance Bot
🏭 预测性维护 Bot
Builds a memory of equipment behavior patterns, failure signatures, and maintenance history. Predicts failures before they happen by comparing current sensor readings to historical pre-failure patterns.
构建设备行为模式、故障特征和维护历史的记忆。通过将当前传感器读数与历史故障前模式比较，在故障发生前预测故障。
  • Failure pattern recognition across equipment fleet
  • 跨设备群的故障模式识别
  • Maintenance ROI tracking over time
  • 随时间追踪维护 ROI
Gaming / Narrative
游戏/叙事
🎮 Persistent World NPC Bot
🎮 持久世界 NPC Bot
NPCs that genuinely remember every player interaction across sessions. Dynamic narrative that reacts to accumulated player choices, not just current session state. World that evolves with player history.
真正记住每个玩家跨会话交互的 NPC。根据积累的玩家选择（而非仅当前会话状态）动态叙事。随玩家历史演进的世界。
  • Cross-session narrative continuity
  • 跨会话叙事连续性
  • NPC memory of past betrayals, favors, alliances
  • NPC 对过往背叛、恩情、联盟的记忆
Marketing
营销
📣 Content Strategy Bot
📣 内容策略 Bot
Tracks which content formats, topics, and distribution channels actually perform for your specific audience over time. Learns from engagement data and proactively suggests the next high-probability content.
随时间追踪哪些内容格式、主题和分发渠道对你的特定受众真正有效。从互动数据学习，主动建议下一个高概率内容。
  • Audience-specific performance memory
  • 针对受众的表现记忆
  • Cross-channel attribution learning
  • 跨渠道归因学习
Infrastructure
基础设施
🛡️ Infrastructure Watchdog Bot
🛡️ 基础设施监控 Bot
Builds baseline behavior models for your entire infrastructure. Detects drift, capacity issues, and configuration anomalies by comparing current state to historical learned norms — not just threshold rules.
为你的整个基础设施构建基线行为模型。通过将当前状态与历史学习规范比较（而非仅依赖阈值规则）检测漂移、容量问题和配置异常。
  • Learned normal vs. threshold-based alerting
  • 学习到的正常值 vs 基于阈值的告警
  • Cross-service correlation intelligence
  • 跨服务关联智能
Multi-Agent
多智能体
🕸️ memMesh — Collective Intelligence
🕸️ memMesh——集体智能
Connect multiple specialised agents (researcher, builder, reviewer) through a shared memU PostgreSQL store. When one agent discovers a fact, all others know immediately. Past failures prevent future regressions automatically.
通过共享的 memU PostgreSQL 存储连接多个专业化智能体（研究者、构建者、审查者）。当一个智能体发现事实时，所有其他智能体立即知晓。过去的失败自动防止未来的回归。
  • Shared memory pool with per-agent access control
  • 带按智能体访问控制的共享记忆池
  • Collective pattern recognition across agent fleet
  • 跨智能体群的集体模式识别
◆ Security安全

Built for enterprise security

为企业安全而生

🏠
100% Local-First
100% 本地优先
All memory, logs, and config on your infrastructure. Nothing to memU servers. Use Ollama for zero external calls — fully air-gapped operation.
所有记忆、日志和配置在你的基础设施上。不向 memU 服务器传输任何内容。使用 Ollama 实现零外部调用——完全气隙运行。
📋
Apache 2.0 — Full Transparency
Apache 2.0——完全透明
All memories are plain Markdown files — readable and auditable by humans at any time. No black boxes. Full source code available. 27 contributors, 528 forks.
所有记忆都是普通 Markdown 文件——随时可由人工读取和审计。没有黑盒。完整源代码可用。27 位贡献者，528 个分叉。
🔒
Multi-Tenant Isolation
多租户隔离
Physical filesystem isolation per user: agent_id/user_id/category.md. Dynamic per-user capacity limits. RBAC access control. One user's activity never affects another's memory.
按用户物理文件系统隔离：agent_id/user_id/category.md。动态按用户容量限制。RBAC 访问控制。一个用户的活动永远不影响另一个用户的记忆。
🗑️
GDPR / Right to Erasure
GDPR / 删除权
Delete at any granularity — individual items, categories, or all data for a user. Full memory export before deletion. You own every byte, always.
任意粒度删除——单个条目、分类或用户的所有数据。删除前完整记忆导出。你始终拥有每个字节。
📦
Offline Capable
离线运行
With Ollama, memU operates with zero external network calls. Fully air-gapped. Ideal for regulated industries, government, and maximum data sovereignty requirements.
使用 Ollama，memU 零外部网络调用运行。完全气隙。非常适合受监管行业、政府和最高数据主权要求。
🔑
Secure Credential Management
安全凭证管理
API keys stored in environment variables, never in memory files. Set API spending limits at provider level. Regular security auditing recommended for production deployments.
API 密钥存储在环境变量中，永不存入记忆文件。在提供者层面设置 API 支出限制。建议对生产部署进行定期安全审计。
◆ Ecosystem生态系统

The memU ecosystem

memU 生态系统

Four repositories make up the full memU stack. Built by NevaMind-AI with 27 contributors, 528 forks, and a growing partner ecosystem.

四个仓库构成完整的 memU 技术栈。由 NevaMind-AI 构建，拥有 27 位贡献者、528 个分叉和不断增长的合作伙伴生态系统。

RepoDescription描述Links
memUCore memory engine — Python 3.13+, Apache 2.0, 12k+ stars核心记忆引擎——Python 3.13+，Apache 2.0，12k+ starGitHub →
memU-serverFastAPI REST backend + Temporal workflow engine + RBACFastAPI REST 后端 + Temporal 工作流引擎 + RBACGitHub →
memU-uiVisual memory dashboard — live category browser + analytics可视化记忆仪表板——实时分类浏览器 + 分析GitHub →
memU-sdk-goGo SDK for memU clients — language-agnostic integration面向 memU 客户端的 Go SDK——语言无关集成GitHub →
🤝
Partners: TEN Framework · OpenAgents · Milvus · xRoute AI · Jaaz · Buddie AI · Bytebase · LazyLLM. Community integrations: n8n · LangGraph · AutoGPT · Dify · LlamaIndex · memMesh. Active PR Hackathons with prize pools — see GitHub for current events.
合作伙伴：TEN Framework · OpenAgents · Milvus · xRoute AI · Jaaz · Buddie AI · Bytebase · LazyLLM。社区集成：n8n · LangGraph · AutoGPT · Dify · LlamaIndex · memMesh。带奖金池的活跃 PR 黑客马拉松——在 GitHub 查看当前活动。
◆ Roadmap路线图

What's coming in 2026

2026 年即将推出的功能

Available (v1.3+)已可用（v1.3+）
  • 4-layer memory architecture (Layer 4 in dev)
  • 四层记忆架构（第四层开发中）
  • Dual RAG + LLM retrieval
  • 双模 RAG + LLM 检索
  • pip install memu-py (Python 3.13+)
  • pip install memu-py（Python 3.13+）
  • PostgreSQL + pgvector backend
  • PostgreSQL + pgvector 后端
  • memU-server + Temporal + UI
  • memU-server + Temporal + UI
  • Go SDK
  • Go SDK
  • OpenRouter + multi-profile LLM
  • OpenRouter + 多配置 LLM
  • Cloud API (memu.so)
  • 云端 API（memu.so）
  • Telegram, Discord, Slack, Feishu
  • Telegram, Discord, Slack, 飞书
In Progress开发中
  • Intention Layer (Layer 4) — cognitive anticipation
  • 意图层（第四层）——认知预判
  • WhatsApp + Email integration
  • WhatsApp + 邮件集成
  • Web UI dashboard
  • Web UI 仪表板
  • AutoGen multi-agent integration
  • AutoGen 多智能体集成
  • Advanced RBAC + SSO
  • 高级 RBAC + SSO
  • Memory analytics dashboard
  • 记忆分析仪表板
Planned计划中
  • End-to-end memory encryption
  • 端到端记忆加密
  • Multi-agent memory orchestration
  • 多智能体记忆编排
  • Custom skill development SDK
  • 自定义技能开发 SDK
  • SOC2 / GDPR compliance reports
  • SOC2 / GDPR 合规报告
  • Sensitive data auto-masking
  • 敏感数据自动遮蔽
  • On-device mobile memory (iOS/Android)
  • 移动端本地记忆（iOS/Android）
◆ FAQ

Frequently asked questions

常见问题

memU (github.com/NevaMind-AI/memU) is the open-source memory algorithm framework — the Python library you integrate into your own agents. Install via pip install memu-py. memU Bot (memu.bot) is a pre-built application using memU as its memory layer — ready-to-deploy agent connecting to Telegram, Discord, Slack, and Feishu out of the box. Think: memU = the engine, memU Bot = a production vehicle built on that engine.

memU（github.com/NevaMind-AI/memU）是开源记忆算法框架——通过 pip install memu-py 安装集成到你自己智能体的 Python 库。memU Bot（memu.bot）是以 memU 为记忆层的预构建应用——开箱即连接 Telegram、Discord、Slack 和飞书的即部署智能体。理解为：memU = 引擎，memU Bot = 基于该引擎构建的生产车辆。

Context windows are finite and expensive. Even a 1M-token window fills up for long-running agents, and every token costs money on every call. memU extracts only meaningful items and injects only the relevant subset per query — typically reducing token usage by ~90%. The agent also gets better answers because the context is clean, focused, and semantically organised rather than raw conversation dumps. Running an always-on agent drops from thousands/month to tens of dollars.

上下文窗口是有限的且昂贵的。即使 100 万 token 的窗口对长期运行的智能体也会填满，每个 token 在每次调用时都需付费。memU 仅提取有意义的条目，每次查询只注入相关子集——通常将 token 使用量减少约 90%。由于上下文干净、专注、语义组织（而非原始对话堆积），智能体也能获得更好的答案。始终在线的智能体成本从每月数千美元降至数十美元。

Yes, fully. Install Ollama, pull a local model: ollama pull llama3.2 or ollama pull qwen2.5. Set memU's base URL to http://localhost:11434/v1. Zero external network calls. For most memory use cases, a 7B–14B parameter model delivers excellent results. Models 70B+ require 48GB+ RAM.

完全支持。安装 Ollama，拉取本地模型：ollama pull llama3.2 或 ollama pull qwen2.5。将 memU 的 base URL 设为 http://localhost:11434/v1。零外部网络调用。对于大多数记忆使用场景，7B–14B 参数模型提供出色效果。70B+ 以上模型需要 48GB+ 内存。

The Intention Layer is the fourth and most ambitious layer in memU's architecture — currently in development by NevaMind-AI. It analyzes patterns across accumulated memories to predict what the user needs before they ask. NevaMind-AI describes this as "cognitive anticipation" — the system acts before a prompt, not in response to one. Expected in 2026. Follow the GitHub repo for release announcements.

意图层是 memU 架构中第四个也是最雄心勃勃的层——目前由 NevaMind-AI 开发。它分析积累记忆中的模式，在用户提问前预测其需要。NevaMind-AI 将此描述为"认知预判"——系统在提示之前行动，而非响应提示。预计 2026 年推出。关注 GitHub 仓库获取发布公告。

memU provides migration tools to import existing mem0 memories. The core API operations (add, search, update, delete) are similar in concept. The main difference is that memU will automatically reorganise your imported memories into its 3-layer hierarchy, deduplicate them, and resolve conflicts — immediately benefiting from improved accuracy and reduced token costs.

memU 提供迁移工具导入现有 mem0 记忆。核心 API 操作（添加、搜索、更新、删除）在概念上类似。主要区别是 memU 将自动将导入的记忆重组到其三层层次中，去重并解决冲突——立即受益于更高的准确性和更低的 Token 成本。

See CONTRIBUTING.md on GitHub. Run make install for dev environment, make check before PRs (Ruff, Black, mypy, deptry). memU runs regular PR Hackathons with cash prize pools — check GitHub for active hackathons. Tracks include: integrations, testing, documentation, and new LLM provider support. Join Discord at discord.gg/memu. Contact info@nevamind.ai for enterprise partnership discussions.

请参阅 GitHub 上的 CONTRIBUTING.md。运行 make install 配置开发环境，PR 前运行 make check（Ruff、Black、mypy、deptry）。memU 定期举办带现金奖金池的 PR 黑客马拉松——在 GitHub 查看活跃的黑客马拉松。赛道包括：集成、测试、文档和新 LLM 提供商支持。在 discord.gg/memu 加入 Discord。联系 info@nevamind.ai 讨论企业合作。