MemOS Cloud OpenClaw Plugin

MemOS Cloud OpenClaw Plugin is an official plugin maintained by the MemTensor (OpenMem) organization, designed for OpenClaw, Moltbot, and ClawDBot Agent frameworks to inject long-term memory capabilities into Agent systems.

The plugin operates through lifecycle hooks: at the before_agent_start stage, it automatically calls the MemOS Cloud memory search API to perform semantic search based on the user prompt, injecting relevant memories into the Agent context as <memories> blocks; at the agent_end stage, it asynchronously writes the current turn's dialogue to MemOS Cloud for continuous memory accumulation.

Key Features#

Memory Recall: Supports global memory search (cross-session recall without conversation_id), configurable recall count (memoryLimitNumber) and relevance threshold (relativity).

Memory Write: Supports asynchronous writes (enabled by default, non-blocking), two capture strategies (last_turn / full_session), and custom tags on write.

Model-based Recall Filtering: Optionally sends recall candidates to a self-hosted OpenAI-compatible model (e.g., Ollama qwen2.5:7b) for secondary relevance filtering.

Multi-Agent Support: Data isolation via multiAgentMode + agent_id, with allowedAgents whitelist and agentOverrides for per-Agent configuration overrides.

Knowledge Base Enhancement: Binds global knowledge base IDs (comma-separated) for domain knowledge augmentation during retrieval.

Operations: Built-in Config UI auto-starts on gateway launch (default http://127.0.0.1:38463) for visual configuration editing.

According to official claims, intelligent memory retrieval can reduce token usage by approximately 72% compared to loading full chat history.

Installation#

Prerequisites: Node.js runtime; OpenClaw / Moltbot / ClawDBot gateway; MemOS Cloud API Key.

NPM Install (recommended):

openclaw plugins install @memtensor/memos-cloud-openclaw-plugin@latest
openclaw gateway restart

Manual Install (Windows workaround): Download .tgz from NPM, extract to local extension directory, configure load.paths in openclaw.json.

Get API Key: Register at MemOS Dashboard and copy the key.

Minimal config (~/.openclaw/.env):

MEMOS_API_KEY="mpg-..."

Key Configuration#

Config priority: plugin config → env files (~/.openclaw/.env → ~/.moltbot/.env → ~/.clawdbot/.env).

Architecture#

• Entry file index.js, core logic in lib/ directory
• Three framework adapter descriptors: openclaw.plugin.json / moltbot.plugin.json / clawdbot.plugin.json
• Recall results injected via appendSystemContext for stable protocol, <memories> blocks in prependContext
• Authentication: Token Auth (Authorization: Token <MEMOS_API_KEY>)
• External APIs: POST /search/memory (recall), POST /add/message (write)

Ecosystem#

• Parent project: MemOS — AI Memory OS with unified memory API, multimodal memory, knowledge base management
• Organization: MemTensor (OpenMem)
• Host frameworks: OpenClaw, Moltbot, ClawDBot
• Recall Filter dependency: User self-hosted OpenAI-compatible models

Unconfirmed Information#

• MemOS Cloud pricing and SLA not documented; Dashboard page not directly verifiable
• Recall Filter model performance requirements unspecified (only qwen2.5:7b example given)
• Whether Cloud Plugin and Local Plugin can coexist is unclear
• memos.memtensor.cn domain suggests China mainland node; other regional endpoints unconfirmed
• Difference between appendSystemContext and prependContext injection methods not detailed
• Full agentOverrides configuration schema not provided