MemTensor/MemOS

MemOS 2.0 – A Memory Operating System for LLMs & AI Agents

What it is – MemOS (Memory OS) is a service that gives large‑language‑model agents a persistent, searchable, and editable memory store. It treats memories as a graph of nodes (text, images, tool traces, personas, etc.) that can be queried with natural‑language or vector search, and it can run either as a fully managed cloud API or as a self‑hosted service on‑premise.

Why it matters – Current LLM deployments treat context as a short‑term window; MemOS adds a long‑term “brain” that lets agents recall past interactions, evolve skills, and share or isolate knowledge across users, projects, or multiple agents. Benchmarks cited in the repo (LoCoMo 88.83, LongMemEval 89.20, OmniMemEval leader) suggest it outperforms existing commercial memory products.

---

Core Features (as described in the README)

Feature	What it does
Unified Memory API	One set of HTTP endpoints to add, retrieve, edit, and delete memories; the data model is a graph rather than a black‑box vector store.
Multi‑modal support	Stores text, images, tool‑trace logs, and persona data together, enabling joint retrieval and reasoning.
Memory cubes	Logical “knowledge bases” that can be composed, isolated, or shared across users/agents.
Asynchronous ingestion (MemScheduler)	Handles high‑concurrency writes with millisecond‑level latency.
Feedback‑driven correction	Memories can be refined via natural‑language feedback (e.g., “replace this fact”).
Hybrid retrieval	Combines full‑text search (FTS5) with vector similarity for fast, accurate look‑ups.
Local‑first SQLite storage	The on‑device plugin stores everything locally, no cloud required.
Cloud service	Hosted version (no infra needed) reachable via an API key.
Agent plugins	Ready‑made plugins for OpenClaw, Hermes, and DeepSeek Harness (both cloud and local variants).
Dashboard / Viewer	Web UI to inspect and edit the memory graph.

---

Typical Use‑Cases

• Personal assistants that remember user preferences (“I like strawberry”) across sessions.
• Customer‑support bots that can retrieve prior tickets or account history.
• Multi‑agent systems where agents share a common knowledge base or keep separate private cubes.
• Skill‑evolution pipelines that store L1 traces, L2 policies, and L3 world‑model updates for continual learning.

---

Quick‑Start Options

Mode	How to start	Infrastructure
Cloud API	Sign up → get an mpg‑… API key → call /add/message and /search/memory endpoints.	None (hosted by MemOS).
Self‑host (Docker)	git clone … && cd MemOS && cp docker/.env.example .env && docker compose up	Neo4j + Qdrant containers (graph + vector DB).
Self‑host (uvicorn)	Same repo, run uvicorn memos.api.server_api:app … after starting Neo4j & Qdrant yourself.	Neo4j + Qdrant services.
Local Plugin	npm install @memtensor/memos-local-plugin (or run the provided install.sh script) inside an existing OpenClaw, Hermes, or DeepSeek Harness environment.	Only Node.js; memory stored in local SQLite.
Cloud Plugin	openclaw plugins install @memtensor/memos-cloud-openclaw-plugin@latest (or the DSH equivalent).	Uses MemOS Cloud; no local DB needed.

---

Example Code (Python, Cloud API)

import requests
API_KEY = "mpg-..."
BASE = "https://memos.memtensor.cn/api/openmem/v1"
HEADERS = {"Authorization": f"Token {API_KEY}", "Content-Type": "application/json"}
# Add a memory
requests.post(f"{BASE}/add/message", headers=HEADERS, json={
    "user_id": "alice",
    "conversation_id": "conv_001",
    "messages": [{"role": "user", "content": "I like strawberry"}],
})
# Query it back
resp = requests.post(f"{BASE}/search/memory", headers=HEADERS, json={
    "query": "What do I like?",
    "user_id": "alice",
})
print(resp.json())

---

Community & Resources

• Docs & tutorials – https://memos-docs.openmem.net/
• GitHub – https://github.com/MemTensor/MemOS (issues, discussions)
• Discord – https://discord.gg/Txbx3gebZR
• Paper – arXiv 2507.03724 (MemOS: A Memory OS for AI System)

---

License

Apache 2.0 (see LICENSE in the repo).

---

Bottom line – MemOS provides a production‑ready, plug‑and‑play memory layer for LLM‑based agents, with both cloud‑managed and fully local deployment paths, and it already ships integrations for several popular open‑source agents.