Workflow Moves Toward Engineering! The Domestic Open-Source AI Agent Runtime Framework MateClaw Releases Major Update v1.5.0

Recently, the open-source AI Agent Runtime framework MateClaw officially released version 1.5.0. This iteration did not blindly pursue the number of models integrated but focused on improving the essential infrastructure required for Agents to operate effectively in real teams. The core was centered around three key dimensions: goal acceptability, knowledge base consistency maintenance, and multi-user memory isolation.

In real business scenarios, traditional agents often rely solely on a vague "completion score" to assess task progress, making it difficult for managers to identify exactly where a task is stuck. To break this black box, MateClaw v1.5.0 introduced a new "Goal Checklist" mechanism. This mechanism allows the system or running large models to dynamically break down a macro goal into multiple independent verification criteria. The built-in evaluator will review each criterion at different stages and record evidence. Only when all criteria on the checklist are passed will the task be marked as completed. This digital checklist design not only makes the end boundary of tasks clearer but also provides precise contextual injection for Agent's automatic follow-up.

As the most significant engineering effort in this update, MateClaw upgraded the traditional knowledge base (LLM Wiki) into a self-maintainable "Knowledge Engine." The new version supports wiki-link functionality similar to Wikipedia, with comprehensive page renaming cascade rewriting, deletion cleanup, and bad link scanning capabilities for consistency maintenance. More impressively, the new system introduced the concept of "knowledge hierarchy," dividing content into basic fact layers and experience summary layers. Once the underlying "fact pages" change, all dependent "experience pages" will be automatically marked as needing review, fundamentally solving the industry-wide issue of "facts changed, conclusions outdated." Additionally, through the PageType Profile feature, administrators can configure structured fields and markdown templates for different page types and implement fail-safe level fine-grained access control based on the matrix of "employee + knowledge base + page type."

To ensure AI agents can safely integrate into multi-user collaboration environments, MateClaw v1.5.0 fully enabled the multi-user memory isolation (Memory per-owner) feature. The system uses owner identifiers (owner_key) and three visibility scopes at the bottom layer to ensure that the same Agent serving multiple different users via Web console, instant messaging (IM) channels, or third-party APIs does not mix personal privacy or long-term memory.

In addition to these three main features, the new version brings a series of stability improvements for production environments. These include supporting binding employees to a default knowledge base for Wiki, optimizing the model selection chain to make preference provider routing truly effective, and integrating a new Claude Opus4.8 model entry. In terms of engineering details, multimedia files generated by tools are now securely stored and automatically cleaned up by scheduled tasks. The default read timeout for the MCP tool has been relaxed to 60 seconds to reduce false positives. Meanwhile, WeChat, Enterprise WeChat, and Feishu channels have fully integrated a unified inbound media pipeline, supporting file type determination using feature codes and exponential backoff retries, significantly improving the reliability of complex file interactions in group chat scenarios. Currently, the upgrade configuration of this version is fully compatible with historical data, and the migration is automatically executed by Flyway, accelerating the transition of AI agents toward true industrial-level productivity deployment.