Xiaomi Open Sources AI Coding Agent With Persistent Memory

0
1
Xiaomi's Open Source Coding Agent Adds Self-Maintaining Memory
Xiaomi's Open Source Coding Agent Adds Self-Maintaining Memory

Xiaomi has open-sourced MiMo Code V0.1.0 under the MIT licence, introducing a persistent-memory AI coding agent designed to maintain context across long software development projects and reduce workflow interruptions.

Xiaomi has released MiMo Code V0.1.0 as an open-source AI coding agent under the MIT licence, giving developers the freedom to use, modify, and extend the software. The release aims to address one of the biggest limitations of AI-powered programming assistants: losing context during long development sessions.

The standout feature of MiMo Code is its persistent memory system. Unlike conventional coding assistants that rely solely on a model’s context window, MiMo Code uses a dedicated background subagent to continuously manage project context. As conversations approach context limits, the system automatically summarises previous interactions, stores structured memory, and preserves continuity throughout extended coding workflows.

To further improve long-term memory management, Xiaomi has introduced a feature called “/dream”. Running automatically every seven days, it launches a maintenance agent that reviews old sessions, removes duplicate memories, verifies file paths, and compresses stored information into an updated long-term memory repository.

Built on the open-source OpenCode project, MiMo Code incorporates Xiaomi-developed memory and workflow enhancements. The tool includes free access to MiMo-V2.5 and can also connect to third-party AI services such as DeepSeek, Kimi, and GLM, helping developers avoid vendor lock-in.

MiMo Code also introduces Compose mode, which enables the agent to handle workflows spanning planning, design, coding, testing, and review from a single goal prompt. Xiaomi additionally developed a dedicated Harness framework optimised for MiMo models.

According to Xiaomi, MiMo Code achieved scores of 62% on SWE-Bench Pro and 73% on Terminal Bench 2, outperforming Claude Code by around five percentage points while using the same base model.

LEAVE A REPLY

Please enter your comment!
Please enter your name here