Xiaomi Open-Sources MiMo Code V0.1: An AI Coding Assistant That Remembers Your Conversations
Xiaomi has open-sourced MiMo Code V0.1, a terminal-based AI coding assistant under MIT license. Its standout features are persistent memory (session-to-session context retention via SQLite FTS5), a Compose orchestration mode for full development lifecycle automation, and voice input — all compatible with every major LLM provider.
💡 What You Will Learn
Xiaomi has open-sourced MiMo Code V0.1, a terminal-based AI coding assistant under MIT license. Its standout features are persistent memory (session-to-session context retention via SQLite FTS5), a Co
Xiaomi Open-Sources MiMo Code V0.1: An AI Coding Assistant That Remembers Your Conversations
Xiaomi has dropped another bombshell.
On June 11, Xiaomi officially released and open-sourced MiMo Code V0.1 — a terminal-based AI coding assistant, MIT licensed, with both code and models fully public.
GitHub hit 1200+ Stars within 24 hours.
This is not another 'Cursor wrapper.' MiMo Code is built on OpenCode, but adds several unique features — cross-session memory, a self-evolving system, and voice input. Plainly put: it is not a forgetful assistant, but a 'coding buddy' that remembers who you are and what your project looks like.
Three Hardcore Upgrades
- Persistent Memory System. This is MiMo Code's biggest differentiator. It implements three-layer memory based on SQLite FTS5: project-level big-picture context (architecture, coding conventions), session-level state snapshots, and task-level progress tracking.
What does this mean? With other tools, every new session starts from scratch — you have to re-explain the background. MiMo Code automatically saves checkpoints, restoring context on next launch so the agent does not have to re-understand your project.
Then there is the /dream command. Every 7 days, an independent agent scans your conversation history, extracting valuable knowledge to update project memory while cleaning out stale entries. It actually gets smarter over time.
- Compose Orchestration Mode. This is not just 'write me a function.' Hit Tab to switch to Compose mode, enter a rough requirement, and it autonomously completes: design to plan to code to test to review to verify to merge — the entire development lifecycle in one pipeline. Each step has a dedicated sub-agent, with an independent judge model to determine whether the task is truly done.
Benchmark results are impressive: on SWE-Bench Pro V2 and Terminal Bench 2, MiMo Code scored 5 percentage points higher than Claude Code under the same model conditions.
- Voice Input. Based on MiMo-V2.5-ASR, speak a command to modify code, issue instructions, or even trigger 'send' and 'execute' — hands-free. A game-changer for live coding and streaming developers.
Ecosystem Ambition: Zero-Cost Migration
MiMo Code's smartest move is its compatibility strategy: works with Claude Code, OpenAI, DeepSeek, Kimi, GLM, and all major model providers; built-in MiMo Auto free tier — zero config to get started; one-click import from Claude Code, retaining existing credentials; full support for OpenCode's LSP, MCP, and plugin ecosystem.
In short, Xiaomi is not trying to compete with Claude Code for users — it wants Claude Code users to seamlessly switch over.
Combined with Xiaomi's Token Plan subscription system (4 tiers from Lite 39 yuan/month to Max 659 yuan/month), the strategy is clear: win developer mindshare with open source, monetize enterprise users with API services.
Honest Impressions
The good. MiMo Code's positioning is spot-on. It is not fighting the IDE war against Cursor — it is carving out the 'terminal AI assistant' niche. Persistent memory + Compose mode genuinely solve real pain points — the AI 'amnesia' problem in large projects finally has a proper fix.
The concerns. First, how long will MiMo-V2.5's free tier last? Reddit users are already complaining about degraded free-tier performance, suspecting a 'bait and switch' strategy. Second, reliance on the OpenCode ecosystem — MiMo Code does not have its own harness. If upstream changes, MiMo Code becomes vulnerable. Third, SWE-Bench score inflation concerns — MiMo-V2.5-Pro scores 57.2% on the harder SWE-Bench Pro, still behind Claude Mythos's 62.3%.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.
