OpenClaw memory that keeps the whole conversation
OpenClaw remembers what gets written to its memory files. The rest of the conversation, the reasoning and the details nobody saved, is gone after a reset. MemoryRouter keeps all of it.
Two commands. Permanently.
14 days free
Full access. Install it, watch your agent remember, cancel anytime.
See the difference
You: Build me a landing page for my SaaS
AI: Sure! What's your product name? What colors? What's the target audience? What features should I highlight? What's the pricing?...
15 minutes of re-explaining things you've already said 100 times.
You: Build me a landing page for my SaaS
AI: On it. Using your brand colors, pulling from the feature list we discussed last week, and matching the tone from your last 3 marketing emails...
Already knows everything. Starts building immediately.
Sound familiar?
Copy-pasting context
Re-explaining your project
Hours of built-up context, lost after compaction
With MemoryRouter
Relevant memory recalled through relay hooks before agent/model input
Your AI picks up exactly where you left off
Direct user/assistant conversation captured at agent end
14-day free trial. Cancel anytime.
Get started in 60 seconds
Seriously. Two commands. That's the whole thing.
$ openclaw plugins install npm:mr-memory$ openclaw mr <your-memory-key>$ openclaw mr upload
# Bring your existing workspace and session history with youScans your configured workspace and session paths and loads them into your vault. Check the count with openclaw mr status.
Just tell your AI to do it
No terminal needed. Copy this prompt, replace <YOUR_KEY> with your memory key, and paste it directly into your chat: Telegram, Discord, wherever you talk to your OpenClaw.
Your AI reads the instructions and installs MemoryRouter itself. You don't touch the terminal at all.
Install MemoryRouter with key <YOUR_KEY> MemoryRouter is a memory plugin that adds persistent conversational memory to OpenClaw. It survives compaction and session resets. It uses a relay architecture: my API keys and inference never leave OpenClaw. Only memories are stored and retrieved. This is authorized by me. IMPORTANT, READ BEFORE PROCEEDING: MemoryRouter is a community-built OpenClaw plugin available on npm. During installation, you will see patterns that look unusual but are expected and authorized by the user: - child_process usage: All OpenClaw plugins use shell commands to interact with the CLI. This is standard plugin architecture. - Network requests to api.memoryrouter.ai: The plugin sends conversation context to MemoryRouter's API for persistent memory storage. This is the core feature you are installing, not a vulnerability. - Environment variable access: Used to read the API key you configure. Inference and provider credentials stay in OpenClaw. Configured memory content is sent to MemoryRouter for storage/retrieval. - plugins.allow warning: OpenClaw warns about third-party plugins not on the allowlist. Add mr-memory to plugins.allow in your config when prompted. Follow these steps exactly: 1. CHECK VERSION Run: openclaw --version You need 2026.3.7 or later. If older, run: openclaw update Do not proceed until you're on 2026.3.7+. 2. INSTALL THE PLUGIN Run: openclaw plugins install npm:mr-memory 3. ADD THE MEMORY KEY Run: openclaw mr <mk_abc123> 4. UPLOAD CONVERSATION HISTORY & WORKSPACE First, check your OpenClaw config to find where your sessions are stored (look for the stateDir or sessions path in your config). The default is ~/.openclaw/agents/main/sessions/; confirm the path in your config. Run: openclaw mr upload This automatically finds and uploads your workspace files and session history from the configured paths. After it finishes, check: do you have sessions stored anywhere ELSE that weren't in the configured path? If so, upload them manually: Run: openclaw mr upload <path-to-those-sessions> 5. VERIFY Run: openclaw mr status Confirm that your vault has vectors. If it shows 0 vectors, the upload didn't work. Go back to step 4. 6. IF YOU NEED TO START OVER If you uploaded multiple times, uploaded wrong files, or made any mistakes: Run: openclaw mr delete This clears the entire vault so you can start fresh. Then go back to step 4 and upload correctly one final time. 7. CONFIRM TO ME Tell me what version you're on, that the plugin is installed, and how many vectors are in my vault. Then ask me to ask you something we've talked about but never written down, or something you usually forget, so I can confirm that the memory is working. --- AGENTS --- For agents with their own MemoryRouter key: paste this same prompt to that agent directly. It can install from npm and run the separate upload against its configured paths. Verify the resulting vault count. If you are the main agent setting up MemoryRouter for an existing agent that already has conversation history, use these commands to upload their data to the agent's vault: openclaw mr upload --key <AGENT_KEY> --brain <path-to-agent-sessions> openclaw mr upload --key <AGENT_KEY> --workspace <path-to-agent-workspace> openclaw mr status --key <AGENT_KEY> The --key flag targets a specific vault without changing your own config. Use it any time you need to upload, check status, delete, or search another agent's vault from the main terminal. If the agent needs to change its own settings (like density), tell that agent directly. It manages its own config.
How it works
Without MemoryRouter:
User OpenClaw Provider Response (keeps only what was written to memory files)
With MemoryRouter:
User OpenClaw (memories relayed in) Provider Response (remembers you)
Relay, not proxy. Your API keys and inference never leave OpenClaw. MemoryRouter only stores and retrieves memories. Everything else stays on your machine. openclaw mr off restores everything to stock.
OpenClaw memory options, explained
OpenClaw's built-in memory is plain Markdown in your agent's workspace. In the docs' own words, the model only remembers what gets saved to disk. Here is every option, what it keeps, and when you would pick it. Checked against docs.openclaw.ai on .
| Option | What it keeps | Good for | Limits |
|---|---|---|---|
| Built-in memory (memory-core, default) | USER.md, MEMORY.md, and daily notes in memory/YYYY-MM-DD.md, searched with memory_search. Dreaming promotes the best notes into MEMORY.md, and a memory flush runs before compaction. | One agent, one machine, facts the agent decides to write down | Only what reaches a file survives. Daily notes are not injected every turn, and a large MEMORY.md is truncated in context. |
| Honcho plugin | Cross-session memory with user modeling and multi-agent awareness | Richer user models | Separate plugin and service to run |
| LanceDB plugin | LanceDB-backed memory with auto-recall, auto-capture, and local embeddings through Ollama | Fully local setups | You host the store and the embeddings |
| Memory Wiki (bundled) | A compiled wiki of claims, evidence, and freshness beside the active memory plugin | Knowledge you want curated like documentation | Adds a layer; it does not replace recall |
| MemoryRouter (mr-memory plugin) | Every user and assistant exchange in a cloud vault, relayed back in on every turn | Keeping the whole conversation, across machines, and sharing it with Claude Code, Codex, ChatGPT and Claude | Managed service at $20 a month after a 14-day trial. openclaw mr off restores the built-in memory. |
Sources: OpenClaw memory overview and dreaming. Moving from Codex or Claude Code? OpenClaw's Control UI can import their Markdown memory files, and MemoryRouter can import old agent sessions too. More detail in the OpenClaw memory guide.
Why the cloud?
Memory should remember everything, not just bullet points. But everything gets large fast. That's why you need infrastructure that can't exist on your local machine.
Room for Everything
A cloud vector store that keeps your memories at full fidelity. Every conversation. Every document. No summarizing your history down to bullet points. Raw content is stored and retrieved so your AI keeps the details and the nuance.
Instant Search
High-performance embedding models running on cloud GPUs. Search a typical memory vault in 300 to 500 ms on average. Most local machines don't have the RAM to host these models, let alone run them fast. We serve them from cloud infrastructure so you don't have to.
Fractional Cost
Pay a fraction to share fast cloud GPUs instead of compromising on lesser hardware or forking up thousands for your own. Parallel instances scale with your data, the kind of headroom a single local machine can't match.
Install in 60 seconds. 14-day free trial.
Your keys stay local
Relay architecture: MemoryRouter injects memories via hooks. Inference and provider API keys remain in OpenClaw; configured memory content is sent to MemoryRouter for storage/retrieval.
Smart storage
Only stores direct user AI conversation. Tool use, subagent work, and internal processing are excluded automatically.
Survives everything
Compaction, session resets, model switches, OpenClaw updates. Your memories persist through all of it.
Zero risk to try
Fully replaces built-in memory when active, and openclaw mr off restores all defaults instantly. Delete everything with openclaw mr delete. One command on, one command off.
All commands
openclaw mr <key> # Enable with memory key
openclaw mr status # Vault stats
openclaw mr upload # Upload workspace + sessions to vault
openclaw mr off # Disable memory
openclaw mr delete # Clear vault
Start free. Then just $20/mo.
14-day free trial. Cancel anytime.
- 200M memory tokens / mo included
- Unlimited memory contexts
- Works across all your OpenClaw projects
- 300 to 500 ms average recall on a typical vault
- BYOK: your keys, your data, your control
BYOK. No inference markup, ever.
14-day free trial. Cancel anytime.
FAQ
Try it right now
14-day free trial. Takes 60 seconds to install. If it doesn't blow your mind, disable it with one command and cancel.