QoderWork launches the "Awareness" feature: allowing AI to learn to remember, reflect, and grow.

On June 16th, Alibaba's desktop AI assistant QoderWork launched the "consciousness" feature, which is the industry's first consciousness system that allows AI assistants to have self-reflection and continuous evolution capabilities. It not only saves conversations but also actively organizes and eliminates redundancies, turning high-frequency tasks into reusable skills. Currently, existing mainstream AI solutions have many drawbacks. Memory systems stack large amounts of information, making it easy to miss key information when retrieving it. The lack of a reflection mechanism prevents high-frequency tasks from automatically solidifying into reusable workflows. These problems make the Agent tool more cumbersome the longer users use it. QoderWork has comprehensively upgraded the memory system and integrated the three modules of memory, reflection, and skill evolution, responsible for "information preservation," "information organization," and "skill production," respectively. In terms of triggering mechanism, QoderWork uses a multi-factor weighted scoring model to comprehensively assess the depth of user intent, conversation length, task complexity, and other dimensions to determine in real-time if each conversation is worth reflecting on. Evolution is only initiated if the score reaches the threshold. QoderWork's reflection mechanism also deeply incorporates the human brain's characteristic of "selective forgetting." It takes a snapshot backup before each reflection, conducts a hard rule post-validation after reflection, and automatically rolls back if it exceeds the threshold to avoid deleting critical information. Additionally, key information always leaves clues, with daily fragmented conversations first sedimenting into short-term memory, and truly valuable content across conversations being promoted to long-term memory. When the AI repeatedly handles a certain type of task, the consciousness system will perceive that "this matter is worth solidifying" and automatically generate a skill suggestion. However, it does not write into the user's skill library on its own; it only offers suggestions, and users can adopt, ignore, or reject them. Rejected suggestions will automatically cool down and not repeatedly disturb the user. According to reports, QoderWork also employs a Fork Session mechanism: each evolutionary task forks out a lightweight sub-session from the main session, reusing the original context, allowing the underlying cache hit rate to reach over 99%. The additional cost of the consciousness system is kept within 5% of the total expenditure of the main conversation. With continuous reflection and evolution, it does not consume too many resources. Unlike most AI memory "black boxes," all memories in QoderWork are stored in open Markdown text locally, allowing users to see, modify, and delete each piece of information. It is reported that QoderWork is currently the most feature-rich desktop AI assistant in China, supporting not only regular chat mode but also launching workbenches for design, presentations, writing, and expert suites for over ten industries such as product, legal, and marketing, meeting the AI needs of users in different professional fields.