#88th of 12 in Memory
MemOS
Memory operating system for agents with cubes, scheduler and hybrid retrieval
Documentation ↗
(opens in a new tab)Website ↗
(opens in a new tab)Repository on GitHub ↗
(opens in a new tab)
- Stars
- 11.8k
- License
- Apache-2.0
- Last commit
- Sep 2026
- Last release
- Sep 2026
- Language
- Python
Overview
MemOS gives LLM apps and agents long-term memory via one API over graph-structured memories, grouped into memory cubes per user, project or agent. Self-hosting runs docker compose for the REST API on port 8000 with Neo4j and Qdrant; a local plugin for OpenClaw, Hermes and DeepSeek Harness instead keeps everything in SQLite with FTS5 and vector search.
Who it is for: Teams building multi-user agent memory with graph and vector stores
Strengths
- Memory cubes isolate or share knowledge across users, projects and agents
- MemScheduler ingests asynchronously for high-concurrency workloads
- Local plugin for OpenClaw, Hermes and DeepSeek Harness is 100% on-device SQLite
- Apache-2.0; two arXiv papers and OmniMemEval benchmark published
Weaknesses
- Self-hosted service requires Neo4j and Qdrant
- Local plugin docs are partly in Chinese; cloud dashboard links go to a cn locale
- LLM, embedder and vector DB keys must be filled in .env before start
- Benchmark table lists scores without comparison baselines in the README
What it needs
- no GPU
- Docker + Compose
- Needs Neo4j, Qdrant
- port 8000
Also in Memory
See all 12| Rank | Project | Score |
|---|---|---|
| 1 | hindsightAgent memory server with retain, recall and reflect operations | 78 out of 100 |
| 2 | MemPalaceLocal verbatim memory for coding agents on ChromaDB with 45 MCP tools | 77 out of 100 |
| 3 | agentmemoryPersistent memory server for coding agents, exposed over MCP and REST | 75 out of 100 |
| #4 | OpenVikingContext database exposing agent memory, knowledge and skills as a filesystem | 74 out of 100 |
| #5 | Mem0Memory layer for agents with a self-hosted server, SDKs and CLI | 73 out of 100 |