NemoClaw's sandbox blocks memory
One command fixes it.

Performance. We're querying A100s on the edge in under 100ms — something local machines can't do. We host the latest top-performing embeddings model, which most local computers don't have enough RAM to host, let alone inference. In the cloud, we spin up multiple parallel instances to handle linear problems exponentially. And we create an infinite-dimensional vector space that can hold and search everything you have at light speed. Local storage hits a ceiling. We don't have one.

QMD rebuilds the entire vector index before every search. The more data you have, the bigger the index, the longer it takes — leading to 30-60 second wait times before your agent even starts responding. Because that search is so expensive, it's avoided and only used when necessary. Sometimes your agent responds instantly, other times it takes forever. The reason? It ran a memory search and you have a lot of data. And because of the size weakness, local memory is forced to take a minimalist approach — consolidating conversations to bullet points. But the detail is what matters. Memory should remember everything, not just the bullet points.

Local isn't free — you either pay with hardware or pay with degraded performance. Running embeddings locally means compromising on model quality, slower search times, and your machine grinding under the load. With MemoryRouter, you fractionally pay for the best hardware in the world — A100 GPUs, top-tier embeddings models, edge infrastructure — and only pay when you use it. It's more efficient to collectively share the best compute than to individually compromise on lesser hardware.

MemoryRouter is a community-built plugin that installs natively via OpenClaw's plugin system. Install with openclaw plugins install mr-memory and you're good to go.

OpenClaw 2026.3.7 or later. That's the version that introduced the hooks the MemoryRouter plugin relies on. Run openclaw --version to check yours, and openclaw update if you need to upgrade.

Nope. The plugin persists across OpenClaw updates. Run openclaw plugins update mr-memory to get the latest version.

Dramatically. MemoryRouter retrieves your memories in under 100ms and injects them before the model even starts. Same model, same prompt — responses go from 30-60 seconds down to 2-3 seconds. We call it OpenClaw Instant.

Your AI knows your name and your config. But does it remember the conversation where you decided to switch frameworks? The bug you spent 3 hours debugging together? The project context you've been building for weeks? MemoryRouter gives your AI persistent memory that compounds over time — every conversation, every decision, every file. It knows YOU. It knows the details. It knows the nuance. It remembers what you said yesterday in exact quotes. It's the difference between starting over every session and picking up exactly where you left off.

No. Our OpenClaw plugin is a relay — memories are retrieved and injected locally inside OpenClaw, before anything reaches your provider. Your API keys, your inference, your data — none of it touches our servers. We only store and retrieve memories — encrypted at rest and in transit.

No. The relay architecture means your API keys and provider calls never leave OpenClaw. We only handle memory storage and retrieval — your inference stays between you and your provider.

Only direct conversations between you and your AI. Tool use, subagent work, and internal processing are excluded automatically.

Yes. openclaw mr delete clears your entire vault instantly. You can also disable memory with openclaw mr off without deleting anything.

MEMORY.md is manual — you have to tell the AI to write things down, and hope it remembers to do it. MemoryRouter is automatic. Every conversation is stored and retrieved intelligently. No maintenance, no forgetting, no hoping.

Yes. Without MemoryRouter, your cron jobs and scheduled tasks run with zero context. With MemoryRouter, they remember everything you've ever discussed. Your morning email triage knows your priorities. Your code review agent knows your architecture. Every automated task runs with full context, never from scratch.

Automatically. MemoryRouter syncs your workspace files on every gateway startup and every new session — using a smart hash manifest that only updates files that actually changed. No rebuilding the entire index every time. Add a file, start a new chat, it's already there. You can also run openclaw mr sync anytime to force a sync.