The Memory Problem Every Vibe Coder Knows Too Well

If you've spent any time working with AI coding agents on substantial projects, you've probably hit the same wall: the longer the session, the more the agent forgets. Context gets compressed, decisions disappear, and suddenly you're re-explaining your project structure for the third time. Xiaomi's new open-source tool, MiMo Code, is betting that persistent memory is the answer—and the company just made it freely available to see if the community agrees.

What MiMo Code Actually Does

MiMo Code is a terminal-native AI coding assistant built on top of the open-source OpenCode agent framework. Xiaomi has extended it with a custom memory architecture that maintains context across sessions using SQLite FTS5 full-text search. The system operates across four distinct layers: project memory stored in a persistent MEMORY.md file, session checkpoints, scratch notes, and per-task progress logs.

The clever part is architectural: MiMo Code deploys a dedicated "checkpoint-writer" subagent that handles documentation while the primary coding agent continues its work. Think of it like having a project architect updating blueprints in real-time while the construction crew builds—everyone stays aligned without pausing the work.

The system also includes a /dream command that periodically reviews historical sessions, deduplicates them, and compresses learnings into long-term memory. There's also a "distill" function that mines past sessions for repeated workflows that can be automated—a feature that echoes approaches recently adopted by OpenAI and Anthropic with their respective models.

The Benchmark Battle

Xiaomi claims impressive numbers. According to their published figures, MiMo Code paired with MiMo-V2.5-Pro outperformed Claude Code paired with Claude Sonnet 4.6 across three standard evaluations:

• SWE-bench Verified: 82% vs. 79%
• SWE-bench Pro: 62% vs. 55%
• Terminal Bench 2: 73% vs. 69%

More interesting is Xiaomi's claim that the harness itself accounts for roughly five percentage points of that gain. Running the same MiMo-V2.5-Pro model in both harnesses showed measurable differences purely from the agent system architecture.

That said, the competitive landscape is more nuanced than Xiaomi's preferred benchmarks suggest. OpenAI's Codex CLI running GPT-5.5 scores 82.2% on the official Terminal-Bench leaderboard—about nine points above MiMo Code's self-reported 73%. However, on SWE-Bench Pro, MiMo Code's claimed 62% does edge above OpenAI's reported 58.6% for GPT-5.5.

Xiaomi notably compares only against Claude Code throughout its materials, avoiding direct comparisons with OpenAI's Codex or Google's Gemini CLI. This selective benchmarking is worth noting when evaluating the claims.

Getting Started

MiMo Code is available now on GitHub under an MIT license. Installation is straightforward:

macOS and Linux:

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

Windows:

npm install -g @mimo-ai/cli

Xiaomi is also offering limited-time free access to MiMo-V2.5, their multimodal flagship model with a million-token context window—no registration required.

Why This Matters for Developer Workflows

The memory problem isn't trivial. For vibe coders working on substantial projects, context loss during long sessions creates friction that undermines the productivity gains AI coding agents promise. The industry has largely treated this as a context window problem—bigger windows, better compression—but Xiaomi's approach suggests a different path: explicit storage-and-retrieval that decides what information matters and when to recall it.

Whether MiMo Code's specific implementation proves superior remains to be seen, but the underlying thesis—that persistent memory architecture deserves serious engineering attention—seems sound. The open-source release gives the community a chance to validate the approach at scale.

For developers tired of re-explaining their projects to AI assistants, this might be worth a look. The million-token context window paired with persistent cross-session memory could genuinely change the vibe coding experience—if the benchmarks hold up under real-world scrutiny.