A coding agent with explicit long-term memory architecture
MiMo Code is a terminal-based coding agent developed by Xiaomi's MiMo team on top of OpenCode, open-sourced under the MIT license. It targets long-horizon automated programming tasks, a category often bottlenecked by context window limits and declining model instruction-following over extended sessions. The core value proposition is an explicit long-term memory architecture that replaces simple context compression with structured storage and retrieval, enabling the agent to maintain high decision quality across dozens or even hundreds of execution steps. This makes MiMo Code particularly suited for developers tackling complex, multi-step coding projects where state continuity is critical.
The fundamental problem MiMo Code solves is that standard coding agents, which pass full conversation history to the model, work for short tasks but fail as turns increase. Two issues emerge: first, the context window fills with tool outputs and logs, forcing compression that weakens distant information; second, the model's instruction-following ability declines with longer inputs, diluting useful constraints. MiMo Code addresses both by moving beyond stateless model calls to a runtime-managed persistence layer that writes structured state to disk and rebuilds context from those records. This matters because real development often requires dozens of steps, and without continuity the agent repeats mistakes or loses track of goals.
The first major feature group is Computation, which focuses on scaling single-turn reasoning. Max Mode generates N candidate solutions in parallel (default 5) at each turn, having each independently complete reasoning and tool-call planning without actual execution. The same model then judges the processes and selects the best plan. This improves performance by 10–20% on SWE-Bench Pro at the cost of roughly 4–5 times token consumption. It addresses the compounding error rate of individual steps in long tasks by investing additional compute for reliability, and when multiple candidates converge, it signals high confidence.
The second major feature group is Memory, which maintains state continuity across multi-turn tasks via a cycle-based architecture. The runtime defines checkpoints far below the context limit (around 20%, 45%, and 70% of budget). At each checkpoint, an independent writer subagent reads the conversation and writes a structured state file with 11 fields to disk without consuming the main agent's attention. When the window nears its limit, a rebuild cuts the current window, opens a new one, and injects the persisted files as a layered prompt kept within roughly 65K tokens. This allows a logical session to extend indefinitely while each physical window remains bounded.
Publisher
A
admin
Launch Date2026-06-16
Pricingfree
Domain Authority
Ahrefs DR80/100