MemU

This note describes MemU, a memory framework built for agents that run all the time. It opens with the definition, then gives self-hosted install steps and two ways to try it, an in-memory test and a PostgreSQL test. The core API section explains the memorize pipeline and the two retrieve modes, one based on retrieval and one driven by a model. Configuration examples follow for OpenRouter and DashScope, and a table lists the cloud API endpoints. The note then lays out the three-layer memory architecture, from resource to item to category, and the proactive features it enables, such as recommendations, email handling and market monitoring. Use cases, tech stack, pros, cons and fit close the page.

Definition

MemU (Memory-U) is a memory framework for proactive AI agents that run 24/7. It helps reduce LLM token costs for always-online agents and lets agents understand user intent without explicit commands.

Installation (Self-Hosted)

pip install -e .

Requirements: Python 3.13+ and an OpenAI API key

Quick Start

Cloud Version

Visit memu.so for the hosted service with 7×24 continuous learning.

Self-Hosted Test

# In-memory mode
export OPENAI_API_KEY=your_api_key
cd tests
python test_inmemory.py

# With PostgreSQL
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_api_key
cd tests
python test_postgres.py

Core APIs

memorize() - Continuous Learning Pipeline

from memu import MemUService

service = MemUService()

# Continuous learning - stores insights, facts, preferences
result = service.memorize(
    user_id="user123",
    content="User prefers concise responses in the morning",
    interaction_type="preference"
)
# Auto-categorizes, extracts skills & knowledge
# Updates user profiles automatically

retrieve() - Dual-Mode Intelligence

# RAG-based retrieval (method="rag")
facts = service.retrieve(
    user_id="user123",
    query="What are user's communication preferences?",
    method="rag"
)

# LLM-based retrieval (method="llm")
context = service.retrieve(
    user_id="user123",
    query="What should I prepare for the user?",
    method="llm"
)

Custom LLM Providers

OpenRouter Integration

from memu import MemUService

service = MemUService(
    llm_profiles={
        "default": {
            "provider": "openrouter",
            "base_url": "https://openrouter.ai",
            "api_key": "your_openrouter_api_key",
            "chat_model": "anthropic/claude-3.5-sonnet",
            "embed_model": "openai/text-embedding-3-small"
        }
    }
)

DashScope (Alibaba)

service = MemUService(
    llm_profiles={
        "default": {
            "base_url": "https://dashscope.aliyuncs.com/compatible-mode/v1",
            "api_key": "your_api_key",
            "chat_model": "qwen3-max",
            "client_backend": "sdk"
        }
    }
)

Cloud API (v3)

Endpoints

Hierarchical Memory Architecture

┌─────────────────────────────────────────┐
│         Category Layer                  │
│   (Summary-level, automatic assembly)   │
├─────────────────────────────────────────┤
│         Item Layer                      │
│   (Targeted fact retrieval)             │
├─────────────────────────────────────────┤
│         Resource Layer                  │
│   (Direct access to original data)      │
└─────────────────────────────────────────┘

Three Layers

1. Resource Layer: Direct access to original data
2. Item Layer: Targeted fact retrieval
3. Category Layer: Summary-level overview + automatic context assembly

Proactive Features

Information Recommendation

• Tracks user interests
• Surfaces relevant content proactively
• Pattern detection for recurring themes

Email Management

• Learns communication patterns
• Drafts responses
• Prioritizes inbox

Trading & Financial Monitoring

• Tracks market context
• Monitors investment behavior
• Alerts on price movements

Use Cases

1. Information Recommendation - Proactively surfaces relevant content based on interests
2. Email Management - Learns patterns, drafts responses, prioritizes inbox
3. Trading & Financial Monitoring - Tracks market context, alerts on price movements

Tech Stack

Pros

Cons

When to Use

• 24/7 AI agents: Always-on assistants that need memory
• User intent understanding: When agents must understand the user without explicit prompts
• Cost-sensitive applications: Production systems that need to optimize token usage
• Recommendation systems: Proactive content/email/trading recommendations
• Personal AI assistants: With hierarchical memory needs

MemU keeps memory for agents that stay online all day. The memorize call stores preferences, facts, and insights, then sorts them and updates the user profile. Retrieve can read stored items or ask a model what to prepare. Three layers hold that memory. Resources keep original data, items keep facts, and categories build a summary. The same store can recommend content, draft email, or watch prices. Production needs PostgreSQL, and the store is not a graph.

References:

• NevaMind-AI/memU
• Cloud API
• Examples

See also: Memory Tool Comparison · Mem0