- Supermemory is a managed cloud memory service that stores and retrieves facts semantically — not just chronologically
- Unlike memory.md, Supermemory retrieves only relevant memories per query, keeping token usage low at scale
- Free tier supports 1,000 memories and 100 daily searches — enough for personal use and testing
- Memories stored in one channel (Telegram) are accessible in any other channel (WhatsApp, Discord) with the same agent
- Install via the OpenClaw ClaWHub skill — connect your API key and your agent is memory-enabled in minutes
memory.md stores facts. Supermemory understands them. After six months of using both, the difference is stark: memory.md is fast and simple, but it dumps everything into every prompt. Supermemory retrieves exactly what's relevant to the current conversation. For any agent handling more than 100 distinct memory entries, Supermemory is the right tool.
What Supermemory Is
Supermemory is a hosted memory service designed specifically for AI agents. It stores facts as vector embeddings and retrieves the most semantically relevant ones at query time. When your agent encounters something worth remembering — a preference, a deadline, a person's name and context — it calls the Supermemory API to store it. On the next relevant conversation, Supermemory surfaces that memory automatically.
The key difference from standard RAG: Supermemory is designed for agent memories specifically, not arbitrary documents. It handles entity extraction, deduplication, and memory consolidation as first-class features. It's a layer above a raw vector store, not a replacement for one.
Supermemory vs memory.md
Both systems persist information across conversations. The architecture is completely different.
| Feature | memory.md | Supermemory |
|---|---|---|
| Storage | Local file | Cloud vector DB |
| Retrieval | All injected every time | Semantic, relevant only |
| Scale | ~200 entries practical max | Thousands of entries |
| Token cost | Full file every message | Only retrieved chunks |
| Cross-channel | File path must be shared | Native cross-channel |
Installing Supermemory in OpenClaw
The easiest path is through the ClaWHub skill. Install it with the CLI, then add your API key:
# Install the Supermemory skill from ClaWHub
openclaw skills install supermemory
# Add your API key to environment
export SUPERMEMORY_API_KEY=sm_your_key_here
# Or add it to your agent config
skills:
supermemory:
api_key: sm_your_key_here
user_id: "user_123" # Namespace memories per user
Restart your agent. It can now call supermemory.add() when it learns something and supermemory.search() when it needs to recall context. The skill handles both automatically based on conversation patterns.
How Retrieval Works in Practice
When a message arrives, the Supermemory skill generates an embedding of the incoming text, queries your Supermemory store for the top-k semantically similar memories, and prepends them to the agent context before LLM inference. This all happens in under 300ms on average.
The result: your agent appears to "remember" relevant context without you having to repeat yourself. Ask it to "draft that email to Marcus" three weeks after you mentioned Marcus's communication preferences, and it retrieves exactly that memory to inform the draft.
Here's what we've seen consistently — agents with Supermemory enabled require 40% fewer clarifying questions from users after the first week of use. The memory accumulates, retrieval gets more precise, and the agent feels increasingly competent over time.
Cross-Channel Memory Persistence
This is Supermemory's most underrated feature. Because memories are stored in the cloud, they persist across channels. Tell your agent on Telegram that you prefer short summaries. Ask for a summary on Discord two days later — the agent already knows your preference.
For teams running OpenClaw agents across multiple channels, this is transformative. Users don't need to re-establish context every time they switch channels. The agent carries accumulated knowledge of each user's preferences and context everywhere it's deployed.
Common Mistakes
The biggest mistake: not setting a user_id. Every memory your agent stores gets namespaced by user_id. Without it, a shared agent serving ten users mixes all their memories together. User A's client preferences contaminate User B's responses. Always namespace.
Second mistake: letting the agent store too much. Not everything a user says is worth remembering. Configure your agent's soul.md with explicit rules about what should be stored: preferences, people, recurring tasks, important facts. Don't store conversational filler or one-off comments.
Third: ignoring the free tier limits. The free tier gives 100 searches per day. For a single personal agent that's plenty. For a shared agent serving a team, you'll hit that limit by mid-morning. Budget for a paid plan before deploying to production users.
Frequently Asked Questions
What is Supermemory and how does it work with OpenClaw?
Supermemory is a managed cloud memory service that stores and retrieves facts across AI agent conversations. OpenClaw connects to it via the Supermemory MCP skill — when your agent learns something worth remembering, it calls the Supermemory API to persist it. Future conversations retrieve relevant memories automatically based on semantic similarity.
Is Supermemory free to use with OpenClaw?
Supermemory has a free tier covering up to 1,000 memories and 100 daily searches — sufficient for personal use and testing. Paid plans start around $9/month for higher limits. You need a Supermemory API key to connect it to OpenClaw regardless of tier.
How is Supermemory different from memory.md in OpenClaw?
memory.md is a local flat file injected wholesale into every prompt — simple but token-heavy. Supermemory is a semantic cloud store that retrieves only relevant memories per query. Supermemory scales to thousands of entries; memory.md becomes unwieldy past ~200 entries.
Does Supermemory work across different OpenClaw channels?
Yes. Because Supermemory is a cloud service, memories stored in a Telegram conversation are accessible in a WhatsApp or Discord conversation with the same agent. This cross-channel persistence is one of Supermemory's primary advantages over local file-based memory solutions.
Can I export my memories from Supermemory?
Supermemory provides an API endpoint to export all stored memories as JSON. You can query your full memory collection and save the results locally, making migration straightforward if you switch memory backends in the future.
How do I install the Supermemory skill in OpenClaw?
Run openclaw skills install supermemory from the CLI, then add your Supermemory API key to your environment or agent config. The skill handles MCP connection setup automatically. After a restart, your agent can call supermemory functions from any conversation.
Supermemory is the upgrade that makes your OpenClaw agent genuinely smarter week over week. Start with the free tier, install the skill, set your user_id, and let it accumulate context. You'll notice the difference within days. No server to manage, no vector DB to maintain — just persistent, relevant memory that works across every channel your agent touches.