---

name: memu description: > Persistent memory infrastructure for 24/7 agents. Replaces flat-file memory with a three-layer architecture (Resource → Memory Item → Memory Category) that reduces token costs by 70-90% and enables proactive context retrieval. Built on NevaMind AI's open-source memU framework (v1.4.0). 92.09% accuracy on LoCoMo benchmark. version: 1.0.0 author: ProjectSnowWork homepage: https://github.com/NevaMind-AI/memU license: AGPL-3.0 tags:

• agent-memory
• long-term-context
• proactive-agents
• token-optimization
• memory-infrastructure
• pgvector metadata: openclaw: requires: env: - OPENAI_API_KEY bins: - python3 primaryEnv: OPENAI_API_KEY

---

memU: Persistent Memory for 24/7 Agents

You are integrating memU, an open-source memory framework by NevaMind AI, into an agent that needs to remember, learn, and act proactively across long-running sessions.

When to Use This Skill

Use memU when the agent needs to:

• Retain and retrieve information across sessions spanning days, weeks, or months
• Reduce token costs from injecting raw conversation history into context (70-90% reduction)
• Act proactively — surface relevant context before the user explicitly asks
• Process multi-modal inputs (conversations, documents, images, logs) into structured memory
• Distinguish between current and outdated information with temporal awareness

Do NOT use memU for:

• Single-session chatbots that don't need persistence
• Simple key-value storage (use a database directly)
• Real-time streaming memory (not yet supported)

Core Concepts

memU organizes memory in three layers:

1. Resources — Raw, immutable inputs (conversations, documents, images). The ground truth.
2. Memory Items — Extracted atomic facts with timestamps, provenance, and confidence scores. Queryable and versioned.
3. Memory Categories — Emergent clusters that self-organize as items accumulate. Enable broad context retrieval.

Two retrieval strategies are available:

• Embedding (RAG) — Vector similarity search. Fast (50-150ms). Best for factual recall.
• LLM — Deep semantic reasoning over memory files. Slower (500-2000ms). Best for nuanced, cross-category queries.

Installation

pip install memu-py

Optional persistent storage:

docker run -d --name memu-postgres \
  -e POSTGRES_USER=postgres -e POSTGRES_PASSWORD=postgres \
  -e POSTGRES_DB=memu -p 5432:5432 pgvector/pgvector:pg16

Minimal Integration

import asyncio
from memu import MemoryService

service = MemoryService(
    llm_profiles={
        "default": {
            "provider": "openai",
            "base_url": "https://api.openai.com/v1",
            "api_key": "sk-your-key",
            "chat_model": "gpt-4o-mini",
            "embed_model": "text-embedding-3-small",
        }
    }
)

# Store
await service.memorize(
    resource_payload=[{"role": "user", "content": "I deploy on Tuesdays."}],
    modality="conversation",
)

# Retrieve
results = await service.retrieve(
    query=[{"role": "user", "content": "When should we deploy?"}],
    method="embedding",
)

Included Files

• README.md — Full architecture explanation, 3 real-world scenarios, cost/performance tables, integration guides, troubleshooting
• FAQ.md — 10 high-frequency questions with detailed answers
• RELEASE.md — Release announcement
• METADATA.yaml — ClawHub form metadata
• examples/ — 4 complete, runnable Python scripts:
  • example_1_minimal.py — In-memory mode, no database
  • example_2_openclaw_integration.py — Replace OpenClaw default memory
  • example_3_production.py — Logging, retries, metrics
  • example_4_scenarios.py — Research assistant, email triage, system monitoring

Attribution

This is a community Skill packaging the official memU project by NevaMind AI. It does not modify or extend memU's code.