MiMoCode's bet is memory, not the model
Most terminal coding assistants forget everything when the session ends. MiMoCode, Xiaomi's terminal-native coding agent, is built around the opposite assumption: that the useful part is what the agent remembers between sessions. It is a fork of OpenCode that keeps the familiar TUI, multi-provider support, LSP, and MCP, then layers on a persistent memory system meant to hold a deep understanding of your project across runs.
You can start it with zero configuration through a free-for-now "MiMo Auto" channel, point it at any mainstream LLM provider, or import your existing Claude Code authentication in one step. Install is a one-line script or npm install -g @mimo-ai/cli, then mimo; the source lives in the XiaomiMiMo/MiMo-Code repository.
The memory and context machinery
The memory system is concrete, not a slogan. Project knowledge lives in a MEMORY.md, structured session state in a checkpoint.md maintained automatically by a checkpoint-writer subagent, scratch notes in notes.md, and per-task logs under tasks/<id>/progress.md. All of it is searchable through SQLite FTS5 full-text search and injected automatically when a session resumes, so the agent does not relearn your project every time.
The context management is the part worth borrowing ideas from. MiMoCode decides when to checkpoint based on the model's context window, and when context nears the limit it rebuilds from the latest checkpoint, project memory, task progress, and recent messages rather than just truncating. A token budget with importance ranking controls how much of that re-enters context. For anyone who has watched a long agent session degrade as it fills its window, this is the right problem to attack.
Two more features stand out. A /goal command sets a stop condition, and when the agent tries to stop, an independent judge model checks whether the condition is actually met, aimed squarely at the "optimistic stop" failure where autonomous agents quit too early. And /dream and /distill mine recent session traces to extract durable knowledge into memory and package repeated manual workflows into reusable skills.
Where it fits, and the caveats
The source is MIT, but use is also subject to a separate use-restrictions file and, for the hosted MiMo services, Xiaomi's terms, so read both before building on it commercially. As a fork tracking OpenCode, MiMoCode inherits that project's strengths and its churn, and the free MiMo Auto channel is explicitly time-limited, so price in a real provider for ongoing use.
If you already live in a terminal coding agent and your frustration is that it forgets, MiMoCode is worth a session or two specifically to see whether the memory and context-reconstruction design holds up on a real project. The model matters less here than the scaffolding around it, which is an unusual and welcome thing to be able to say about a coding assistant.
GitHub: https://github.com/XiaomiMiMo/MiMo-Code
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (0)