Microsoft and Xiaomi Open-Source New Architectures to Fix AI Agent Amnesia

The promise of AI agents—autonomous digital workers that can plan, code, and execute multi-step workflows—has long been bottlenecked by two fundamental flaws: they are difficult to instruct, and they suffer from severe amnesia.[1][2]

When developers want an AI agent to perform a specific task, they typically write a set of instructions in a plain text file, often called a skill document or AGENTS.md file. But optimizing these text files has historically been a manual, trial-and-error guessing game, forcing engineers to constantly tweak sentences to see what the AI responds to best.[4][5]

Furthermore, as an agent works through a complex project, its "context window"—the active memory of the conversation—fills up. Once that limit is reached, the agent begins to forget earlier decisions, file paths, and architectural rules, forcing developers to constantly re-explain the project parameters.[2][3]

This week, two major open-source releases offered a new blueprint for solving both problems without requiring massive, expensive model retraining. Microsoft introduced SkillOpt, a framework that algorithmically perfects agent instructions, while Xiaomi released MiMo Code, a coding assistant with a persistent, long-term memory system.[1][2]

Microsoft Research’s SkillOpt, released under the permissive MIT license, introduces a concept called "text-space optimization." Instead of fine-tuning the billions of mathematical weights inside a large language model, SkillOpt treats the agent's plain-text skill document as the trainable parameter.[1][4]

The framework borrows the rigorous discipline of deep learning—epochs, batch sizes, and validation gates—but applies it entirely to natural language. It works through an iterative loop: a frozen target model attempts a batch of tasks using the current skill document, generating a trajectory of successes and failures.[4][7]

A separate optimizer model then reviews this trajectory, reflects on where the agent went wrong, and proposes bounded edits to the markdown file—adding a constraint, deleting a confusing heuristic, or replacing an example.[4][7]

Crucially, an edit is only accepted if it strictly improves the agent's performance on a held-out validation dataset. The result is a highly refined, compact text file—typically between 300 and 2,000 tokens—that can be checked into version control and shared across teams.[4][7]

The performance gains are substantial. Across various industry benchmarks, SkillOpt boosted the average accuracy of GPT-5.5 by 23.5 percentage points, all without adding a single inference call or compute cost at deployment. Because the skill is just text, it can be transferred seamlessly between different models, from OpenAI's Codex to Anthropic's Claude.[1][4][7]

But perfect instructions are only half the battle if the agent forgets them halfway through a task. This is the exact problem Xiaomi set out to solve with MiMo Code V0.1.0, a new terminal-native AI coding assistant that operates directly in a developer's environment.[2][3]

Xiaomi’s MiMo AI team recognized that simply expanding the context window—even to the one million tokens supported by their bundled MiMo-V2.5 model—only delays the amnesia problem. Instead, MiMo Code introduces an explicit storage-and-retrieval architecture powered by a SQLite full-text search database.[2][6]

As a developer works, MiMo Code employs a dedicated background subagent that continuously monitors the active context. Before the window fills up, this subagent automatically condenses the conversation into structured summaries and writes them to persistent project memory files.[3][6]

This allows the primary coding agent to maintain its operational momentum across sessions that span days or weeks. In internal benchmarks, this architecture allowed MiMo Code to outperform Anthropic's Claude Code on ultra-long, multi-step tasks requiring more than 200 distinct actions.[2][6]

To prevent this external memory from becoming bloated or stale, Xiaomi engineered a self-maintenance mechanism. A specialized command, which runs automatically every seven days, launches a separate maintenance agent to review historical sessions, deduplicate overlapping instructions, verify file paths, and compress the data into an updated long-term memory store.[2][3][6]

Both Microsoft’s and Xiaomi’s releases highlight a broader industry shift toward "progressive disclosure" in AI architecture. Rather than stuffing every possible instruction and piece of context into a single massive prompt, modern agents load specialized knowledge only when needed.[5]

Platforms like GitBook are already adapting to this shift, allowing teams to host version-controlled skill files directly within their product documentation, ensuring that AI agents always pull the most accurate, up-to-date operational guidance.[5]

By open-sourcing these frameworks, Microsoft and Xiaomi are democratizing advanced agentic capabilities. Developers are no longer locked into proprietary, black-box ecosystems; they can now build highly capable, long-memory agents using interchangeable, open-weight models.[1][6]

Ultimately, these innovations signal a maturation in how the software industry treats AI. We are moving away from treating models as omniscient oracles that must be endlessly retrained, and toward treating them as computing engines that simply need better software architecture—reliable instructions and persistent memory—to do their jobs.[1][2]