MemOS

• CollectionMemory & State
• Classificationagent-native
• LicenseApache 2.0
• MCPAvailable
• Agent SkillAvailable
• InterfacesAgent Skill · MCP · REST · SDK · STDIO · http · API

• Website ↗
• Docs ↗
• Official repo ↗
• Full dossier on GitHub ↗

What it does

MemOS is a memory operating system — it manages memory as a system resource for AI agents the way an OS manages memory for processes. Rather than treating agent memory as a simple key-value store or vector database, MemOS unifies three fundamentally different types of memory under one API:

• Parametric memory — knowledge embedded in model weights
• Activation memory — precomputed KV-caches and hidden states for fast context retrieval
• Plaintext memory — searchable, editable text, documents, and graph nodes

The basic unit is the MemCube: a composable memory object with content, provenance, versioning, and lifecycle metadata. MemCubes can be composed, migrated, and fused as the agent accumulates knowledge.

MemOS v2.0 benchmark (March 2026): +43.70% accuracy vs. OpenAI Memory, 35.24% token savings.

Agent-facing interfaces

• Interaction pattern: SDK / REST
• Catalog protocols: Agent Skill, MCP, REST, SDK, STDIO, http, API
• MCP: ✅ Available
• Agent Skill: ✅ Official OpenClaw plugin available
• Onboarding type: sdk

How agents onboard

Interaction pattern: SDK / REST

pip install memos-core

# Configure and start MemOS
from memos import MemOS
memory = MemOS(config="~/.memos/config.yaml")

# Add a memory from agent session
memory.add("User prefers Python over TypeScript for new projects")

# Retrieve relevant memories before agent responds
results = memory.get("user coding preferences")

See the full onboarding section in the dossier.

Protocol surface