Your Coding Agent Has Goldfish Memory. Xiaomi's MiMo Code Just Open-Sourced a Hippocampus.

Your Coding Agent Has Goldfish Memory. Xiaomi's MiMo Code Just Open-Sourced a Hippocampus.

I've watched Claude Code choke on a 400-file refactor because it forgot why we started by step 150. Context windows collapse like cheap scaffolding under wet snow. It's like hiring a brilliant contractor who burns the blueprints every morning and asks you to re-explain the kitchen layout before every swing of the hammer.

Xiaomi's MiMo Code V0.1.0, open-sourced last week, is MIT-licensed and installs with a single terminal command. It doesn't just bring a cheaper model. It grafts a hippocampus onto the OpenCode skeleton: SQLite FTS5, four-layer memory, and a background "checkpoint-writer" subagent that takes notes without holding its breath.

The four layers are brutalist and practical. Project memory lives in MEMORY.md. Session checkpoints and scratch notes catch transient context. Per-task progress logs keep the long arc visible. Every seven days, the /dream command runs like a night-shift janitor—reviewing sessions, deduplicating noise, compressing signal into long-term memory. The distill function is the real killer: it mines your history for repeated workflows and turns muscle memory into executable runbooks.

We've seen memory layers before. Hermes Agent experimented with persistent scratchpads and episodic recall, but it always felt like taping a notebook to a race car—clunky, aerodynamically suicidal, and prone to flying off at speed. Xiaomi's architecture is integrated: the checkpoint-writer is a dedicated subagent, not an afterthought. It journals in parallel so the main agent never breaks flow. That's the difference between a diary duct-taped to your dash and an actual nervous system.

Now, the numbers—because you'll ask.

Benchmark	MiMo Code + MiMo-V2.5-Pro	Claude Code + Sonnet 4.6	Context
SWE-bench Verified	82%	79%	Self-reported
SWE-bench Pro	62%	55%	GPT-5.5 scores 58.6%
Terminal Bench 2	73%	69%	Codex CLI: 82.2% (official)
Human A/B (>200 steps)	>65% win rate	<35%	576 devs, 474 private repos

The scaffolding gain is real. Run the same model through both harnesses and MiMo's workflow engine adds roughly five points. On long-horizon tasks—200+ steps—MiMo Code won over 65% of 1,213 head-to-head pairs across 576 developers and 474 private repos. Under 200 steps, it's a coin flip. That trajectory matters: the harder the task, the more memory pays dividends.

But I'm not going to launder Xiaomi's press release. These benchmarks are self-reported. MiMo Code's 73% on Terminal Bench 2 isn't on the official leaderboard