This is week 18 of my “I Used It for a Week” series. Last week I reviewed Poolside, the coding agent with RLCEF training. This week: the free CLI agent that remembers what you did last week.

MiMo Code is the tool I didn’t expect to like. It’s made by Xiaomi, a Chinese phone company. It ships with a free 1-trillion parameter model. And it has a memory system that no other coding agent offers.

After a week of using it on real projects, I think it’s the most underrated free coding tool out there. The memory feature alone is worth the install.

How It Works

Installation is one command on Mac and Linux:

curl -fsSL https://mimo-code.xiaomi.com/install.sh | sh

On Windows:

npm install -g @xiaomi/mimo-code

Then point it at your project and start talking:

cd my-project
mimo

It drops you into an interactive terminal session. You type what you want, it reads your files, makes changes, and runs tests.

The default model is MiMo V2.5 Pro: 1 trillion parameters, 1 million token context, multimodal (text, code, images). And it’s free.

Day 1: First Impressions

I opened a TypeScript project and asked: “Convert all REST endpoints to use Zod validation schemas.”

MiMo Code read every route file, identified the endpoints, generated Zod schemas that matched my existing types, added validation middleware, and updated the error handling. Then it ran the test suite.

Total time: about 4 minutes. Similar to what Claude Code did in my earlier test.

But here’s where it got interesting. The next day, I came back to the same project and asked: “Add rate limiting to the API endpoints.”

Without me telling it anything about the project, MiMo Code already knew:

• What validation schemas were in place (from yesterday)
• The file structure and naming conventions
• Which endpoints were public vs authenticated
• The tech stack and dependencies

This is the persistent memory system at work. Claude Code would have made me re-explain everything.

Day 2-3: The Memory Feature

The memory system uses SQLite with full-text search. Every session, MiMo Code stores:

• Project structure and file relationships
• Architecture decisions you’ve made
• Code patterns and conventions
• Previous debugging context
• File paths and their purposes

When you start a new session, it loads this context automatically. You can also query the memory directly:

/mem search "rate limiting"

It returns all relevant context from previous sessions. This is genuinely useful for long-running projects where you need continuity.

I tested it by forgetting (intentionally) what I’d worked on for 3 days, then coming back and asking MiMo Code to continue where I left off. It remembered the progress, the decisions, and the remaining TODOs.

Day 4-5: The /dream Command

Every 7 days, MiMo Code can run its /dream command. This is a maintenance agent that:

• Reviews all recent sessions
• Removes duplicate or outdated information
• Verifies file paths still exist
• Compresses memory into long-term storage
• Optimizes the database for faster retrieval

Think of it as garbage collection for your AI agent’s brain. Without maintenance, memory fills up with stale references. MiMo Code handles this automatically.

I ran /dream after 5 days of heavy use. It cleaned up about 30% of the memory entries, removing references to files I’d deleted and consolidating duplicate patterns.

What Blew Me Away

Persistent memory

This is the killer feature. No other coding agent remembers your project across sessions without manual configuration. Claude Code uses static CLAUDE.md files. Cursor has workspace memory but it’s limited. MiMo Code’s memory is dynamic, searchable, and automatic.

Free model

MiMo V2.5 Pro is genuinely good. 1 trillion parameters, 1M context window, and no usage limits announced. For a free tool, the quality is impressive.

Voice control

You can speak commands instead of typing them. This sounds gimmicky but it’s actually useful when your hands are busy or you’re debugging something that requires your full attention.

Open source

MIT license. You can fork it, modify it, self-host it. The code is auditable. For companies with security requirements, this matters.

What Frustrated Me

Xiaomi’s servers

By default, your code goes to Xiaomi’s servers for processing. The agent is open source, but the model calls are not. For proprietary code, this is a security consideration. You can use alternative backends (DeepSeek, Kimi, GLM), but that sacrifices the free MiMo V2.5 Pro model.

Documentation is thin

The docs are sparse compared to Claude Code or Cursor. Figuring out advanced features took trial and error. The community is growing but not as mature as Anthropic’s ecosystem.

Terminal-only

Like Aider, Claude Code, and Poolside, MiMo Code is terminal-only. No inline completions, no visual editor, no IDE integration. For quick edits, I still reached for Cursor.

Occasional hallucinations

The model occasionally generates code that looks right but doesn’t work. This happens less than with cheaper models, but it’s more frequent than with Claude Sonnet 5. Always run the tests.

MiMo Code vs Claude Code vs Poolside

• Claude Code: Best ecosystem, best documentation, most polished UX. But costs $20+/month and has no persistent memory.
• Poolside: Best training (RLCEF), open weights, thinking mode. But no memory system and thinner docs.
• MiMo Code: Best memory, free model, voice control. But Xiaomi servers and less mature ecosystem.

They’re complementary. My ideal setup: Cursor as my editor, Claude Code for complex tasks, MiMo Code for long-running projects where memory persistence helps.

Would I Keep Paying?

It’s free. The model is free. The tool is free. The only cost is sending your code to Xiaomi’s servers (or using alternative backends).

For personal projects and open-source work, the value proposition is unbeatable. For proprietary code, consider the security implications.

Rating: 8/10 — The memory feature is genuinely innovative and the free model is surprisingly capable. The Xiaomi server dependency and thin documentation hold it back from a 9.

FAQ

Is MiMo Code really free?

Yes. MiMo Code is open source (MIT license) and MiMo V2.5 Pro (the default model) is free with no announced usage caps. This could change, but right now there’s zero cost.

Is my code safe with MiMo Code?

Your code is sent to Xiaomi’s servers by default. The agent is open source and auditable, but model calls go to Xiaomi. For proprietary code, consider using alternative backends like DeepSeek or running local models.

How does the memory compare to Claude Code’s CLAUDE.md?

Claude Code’s CLAUDE.md is static and manually maintained. MiMo Code’s memory is dynamic, automatic, and searchable. MiMo Code remembers patterns and decisions across sessions without you maintaining anything.

Can I use MiMo Code with other models?

Yes. MiMo Code supports DeepSeek, Kimi, and GLM as alternative backends. You configure this in the settings.

Does MiMo Code work on Windows?

Yes. Install via npm: npm install -g @xiaomi/mimo-code. Mac and Linux use a curl one-liner.

Related: MiMo Code Complete Guide · MiMo Code vs Claude Code · Best Open-Weight Coding Models 2026

Next week: I Used ZCode for a Week — Z.ai’s desktop agent with Telegram remote control.