The Memory Problem No One Fixed — Until Now
Every AI coding agent resets to zero when you close the terminal session. That context loss is the reason agents that work brilliantly on a single-session task collapse on week-long feature builds. Xiaomi's MiMo Code was built explicitly to solve this.
Released on June 10, 2026 as a fork of OpenCode, MiMo Code adds a persistent memory system that preserves project understanding across sessions and a self-evolution (dream/distill) mechanism that automatically reviews, condenses, and improves stored memory over time. The result is an agent that can carry a multi-session refactor without losing the thread.
Architecture: Computation, Memory, Evolution
The official technical blog describes MiMo Code's design around three themes. Computation: a goal-driven autonomous loop runs on top of a main-loop state machine, with subagent orchestration for parallel sub-tasks. Memory: structured encoding persists project knowledge between sessions so the agent stays coherent across hundreds of steps and multiple logins. Evolution: the dream/distill mechanism periodically reviews and condenses stored context, pruning noise and reinforcing what matters — it learns from each session.
Feature Comparison
| Feature | MiMo Code | Claude Code | GitHub Copilot |
|---|---|---|---|
| Cross-session memory | ✅ Persistent | ❌ Resets | ❌ Resets |
| Open source | ✅ MIT | ❌ Proprietary | ❌ Proprietary |
| Built-in free model | ✅ MiMo Auto | ❌ | ❌ |
| Long-horizon (200+ steps) | ✅ Designed for it | Limited | Limited |
| Subagent orchestration | ✅ | ✅ | ❌ |
| Voice input | ✅ (login required) | ❌ | ❌ |
Run
curl -fsSL https://mimo.xiaomi.com/install | bash or npm install -g @mimo-ai/cli. On first launch, the TUI guides model configuration. Existing Claude Code credentials can be imported in one step — no manual API key setup required.
Model Flexibility: The Ecosystem Bet
MiMo Code is not locked to Xiaomi's models. It ships with MiMo Auto — a free-for-limited-time multimodal model based on MiMo-V2.5 with a 1 million token context window — but supports Xiaomi MiMo Platform OAuth, imported Claude Code config, or any OpenAI-compatible API endpoint. The memory system and agentic workflows are the differentiator. The model layer is pluggable.
Compose mode targets professional-grade delivery from idea to shippable product, framing it as "a professional development team for one person." Real-time streaming voice input via TenVAD and MiMo ASR is available for logged-in users.
With over 5,700 GitHub stars in two days, developers are paying attention. The benchmark claim — outperforming Claude Code on 200-step tasks — is notable, but the more durable differentiator may simply be that it remembers where you left off.
- MIT-licensed, terminal-native open-source coding agent built as a fork of OpenCode
- Persistent cross-session memory keeps project context alive across sessions and restarts
- Self-evolution (dream/distill) automatically condenses and improves stored memory
- Claims to outperform Claude Code on software engineering tasks with 200+ sequential steps
- Ships with free MiMo Auto model (1M token context); supports any OpenAI-compatible API
— MiMo Code Official GitHub Repository (MIT License, v0.1.0)
— v0.1.0 Release Notes and Download
— README: Installation, Quick Start, and Configuration Guide