Microsoft and Xiaomi Open-Source Breakthroughs in AI Coding Memory and Skills

The promise of AI coding assistants has hit a wall: long-horizon tasks. Once a project stretches past 200 steps, models suffer from "amnesia," forgetting earlier decisions as their context windows fill up. Meanwhile, the instructions guiding these agents—known as skills—remain brittle, requiring human developers to play a guessing game of prompt engineering.[1][2][3][5]

But a fundamental shift is underway in how the software industry approaches artificial intelligence. Instead of waiting for the next trillion-parameter model to solve these issues, researchers are building smarter "scaffolding" around existing, frozen models.[3][7]

In a span of just two weeks in mid-2026, two major open-source releases have fundamentally altered the AI engineering landscape. Microsoft introduced SkillOpt, a framework that trains text-based agent instructions like neural networks, while Xiaomi launched MiMo Code V0.1.0, a terminal-native coding harness that gives AI persistent, cross-session memory.[1][2]

Microsoft’s SkillOpt, released under an MIT license, tackles the brittleness of agent instructions. In modern AI applications, agents rely on a folder of markdown files that dictate their behavior, tool use, and constraints.[1][4]

Traditionally, optimizing these markdown files meant a human developer manually tweaking text and hoping for better results. SkillOpt automates this by treating the plain text document as a trainable parameter, shifting the focus from training model weights to training behaviors.[1][3][4]

Using deep-learning concepts adapted for text—such as epochs, batch sizes, and validation gates—a separate optimizer model runs the target agent on a batch of tasks, reflects on its failures, and proposes bounded edits to the markdown file. An edit is only kept if it strictly improves the agent's score on a held-out validation set.[3][4][7]

The results are striking. On OpenAI's GPT-5.5, SkillOpt lifted average accuracy by 23.5 percentage points across industry benchmarks, with zero added inference cost at deployment. Because the final artifact is just an optimized text file, it can be easily shared and transferred across different models and execution environments.[4][7]

While Microsoft solved the instruction problem, Xiaomi’s MiMo Code targets the memory bottleneck. Released as an open-source fork of the OpenCode project, MiMo Code is a terminal-based AI assistant designed specifically for ultra-long, multi-step coding tasks.[2][6]

To cure context-window amnesia, Xiaomi engineered a persistent memory system. Instead of relying solely on the active context window, MiMo Code deploys a dedicated background subagent that continuously monitors the session.[2][8]

As the conversation approaches its token limit, this subagent automatically condenses the interactions into a structured summary, storing it in a persistent memory file. This allows the main coding agent to continue working uninterrupted without losing its grasp on the project's architecture.[2][5][8]

Xiaomi also introduced a novel maintenance feature called "/dream". Running automatically every seven days, this background process reviews old sessions, removes duplicate memories, verifies file paths, and compresses the stored information into an updated long-term repository.[2][5]

This explicit storage-and-retrieval mechanism has yielded impressive benchmark victories. According to Xiaomi, MiMo Code achieved a 62 percent score on SWE-Bench Pro and 73 percent on Terminal Bench 2, outperforming Anthropic’s Claude Code by roughly five percentage points while using the same underlying base model.[2][6]

Both releases underscore a growing consensus in the AI community: the highest-leverage optimization target is no longer the model weights, but the behavioral scaffolding surrounding them. By open-sourcing these tools, Microsoft and Xiaomi are democratizing capabilities that were previously locked behind proprietary enterprise platforms.[1][2][3][7]

Developers can now equip open-weight models like Qwen or DeepSeek with the memory and precision of frontier models, entirely avoiding vendor lock-in. As the focus shifts from training massive neural networks to optimizing agent behaviors, the barrier to building industrial-grade AI software continues to fall rapidly.[2][4][3][8]