MyContext – In-Depth Review of Alibaba Qwen Office's Open-Source Agent Context Infrastructure

1. What is MyContext

MyContext is an open-source Agent context infrastructure introduced by the Alibaba Cloud Qwen Office team. It employs a local-first architecture to automatically process multi-source work data from instant messaging tools such as DingTalk and Feishu, as well as document and meeting systems, into continuously updated and traceable personal work archives. These archives include scope of responsibilities, collaboration relationships, and work conclusions. When encountering conflicting information, semantic analysis is used to determine whether to merge or overwrite it; when uncertainty arises, the system prompts the user for confirmation. Based on these archives, personalized Agents can be built to enable work information queries and human-like automated responses, addressing the long-term memory gap that Agents face in office scenarios.

Image source: Official article
Image source: official article

Technical positioning and domain: MyContext belongs to the Agent context infrastructure domain, focusing on long-term memory and knowledge management in office scenarios. It differs from general-purpose memory layers or RAG systems by deeply integrating with workflows, transforming unstructured information scattered across IM, documents, and meetings into structured work archives that can be accessed by Agents. This fills the gap in the memory foundation for Agent deployment in office environments. Its positioning lies between a data platform and AI Agents, emphasizing the concept of "Context as a Service."

Development background: Developed and open-sourced by the Alibaba Cloud Qwen Office team, leveraging Alibaba's deep experience in office collaboration (DingTalk). The team aimed to address the pain points of fragmented cross-application work data and the lack of long-term context in AI assistants, promoting the practical application of Agents in real work scenarios. The open-source strategy lowers the barriers for enterprise trials and secondary development, accelerating technical validation and community feedback.

Core value: The core value of MyContext lies in transforming scattered work data into continuously updated "work memory," enabling Agents to understand users' responsibilities, collaboration relationships, and historical conclusions, thereby delivering accurate Q&A and human-like automated responses. Additionally, through conflict resolution of information and a fully traceable mechanism throughout the process, it significantly enhances the reliability and security of AI outputs. It solves the problem that traditional RAG systems are unable to handle dynamic workflows and factual conflicts.

Technical features: MyContext adopts a local-first data storage strategy to ensure enterprise data privacy; it includes a built-in semantic analysis engine to handle information conflicts and introduces a manual confirmation step to prevent erroneous overwrites; it supports a hybrid retrieval architecture combining full-text search, semantic recall, and graph query to achieve efficient and accurate information localization; it provides a complete nine-tier architecture, forming a closed loop from data ingestion, standardization, index building, archive generation, to Agent interaction, with each layer capable of independent expansion.

2. Key Features

• Multi-source Data Integration and Standardization: Supports integration with mainstream office IM systems such as DingTalk and Feishu, as well as multi-source data including documents and meeting records. The system automatically performs data retrieval, format standardization, and content parsing, converting unstructured information into a unified internal representation, providing high-quality input for subsequent archival construction. This process supports incremental synchronization to avoid redundant processing.

• Automatic Work Archiving: Based on standardized data, MyContext automatically builds personal work archives, covering areas such as job responsibilities, collaboration relationships, behavioral patterns, and discussion conclusions. Archives are continuously updated incrementally, without requiring manual maintenance from users, and remain synchronized with the latest work updates. Archives are structured around a memory graph, supporting entity relationship queries.

• Intelligent Conflict Resolution for Information: For conflicting information before and after, the system uses semantic analysis to determine whether it should be merged and supplemented or overwritten. When the model cannot determine the correct action, the decision is handed over to the user for manual judgment, and this judgment will not be automatically overwritten by the model, fundamentally preventing the issue of old facts being erroneously overwritten in traditional solutions. This mechanism significantly improves the long-term consistency of the archives.

• Fully Traceable Mechanism: Every piece of information in the archives is labeled with its source (specific message/document), timestamp, and original content, supporting one-click navigation to the source. In Q&A scenarios, the Agent's answers also include source citations, allowing users to verify them at any time, effectively mitigating the hallucination problem in large models. Traceability is a core requirement for enterprise-level applications.

• Hybrid Retrieval and Q&A: Combines three retrieval methods: full-text search (keyword matching), semantic retrieval (vector similarity), and graph query (entity relationships). Users can query work history using natural language (e.g., "What was the plan decided on last week's meeting"), and the system returns precise answers with source citations. The hybrid retrieval architecture achieves a good balance between coverage and accuracy.

• Digital Avatar and Human-like Responses: Based on the work archives, a dedicated Agent is built. This Agent can learn the user's communication style and common expressions, drafting work messages in the user's own voice. Actions with consequences, such as sending messages or modifying documents, require user authorization to ensure security and controllability. The quality of the digital avatar's responses directly depends on the completeness and accuracy of the archives.

• Human-Machine Safety Gate Mechanism: For operations that may have real-world impacts (such as sending messages or modifying documents), the system requires explicit user authorization before execution. This design balances automation efficiency with security risks, preventing the Agent from taking inappropriate actions due to a misunderstanding of the context. The safety gate mechanism is a critical safeguard for Agent deployment.

3. How to Use

1. Environment Requirements and Deployment: MyContext runs as a desktop application and supports Windows 10/11, macOS 12+, and major Linux distributions. Users need to download the latest release package from the GitHub repository (https://github.com/openTrinity/mycontext), extract it, and run it directly—no need to install Python or configure complex dependencies. It is recommended that the device has at least 8GB of memory and 50GB of disk space. If using local AI models, it is advisable to equip the device with an NVIDIA GPU.

2. Authorization and Data Source Access: After the first launch, follow the wizard to complete OAuth authorization for IM accounts such as DingTalk and Feishu. The system requests permissions to read message history, document lists, and meeting summaries. Users can choose the data scope according to their needs. The authorization process uses standard protocols, and credentials are only used for a one-time token exchange; passwords are not stored.

3. Import and Synchronize Data: Once authorization is completed, MyContext automatically pulls historical messages, documents, and meeting records and performs incremental synchronization. Users can also manually specify a particular time period or file directory for import. Data standardization is completed locally, and all data is stored only on the user's local machine, without being uploaded to the cloud.

4. Generate and View Work Profiles: After data import is complete, the system automatically analyzes and aggregates the data to generate a personal work profile. Users can view an overview of the profile in the application interface, including role tags, collaboration relationship graphs, and recent discussion conclusions. The profile will be continuously updated with new data, and users can manually trigger a reindex.

5. Build a Custom Agent: Based on the generated profile, users can create a personal Agent with one click. The Agent supports natural language conversations and can query work-related information, summarize knowledge, or assist in replying to messages. Users can customize the Agent's response style and authorization scope, such as setting conditions for automatic replies.

6. Manual Confirmation for Critical Actions: When the Agent encounters conflicting information or needs to perform actions such as sending a message, the system will pop up a confirmation window for the user to review and execute. Any modifications made by the user to the profile will be locked and will not be overwritten by the model, ensuring the authority of human decisions.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations:

For professionals or teams seeking out-of-the-box solutions with deep integration into office environments, MyContext is the optimal choice. It seamlessly connects with the DingTalk/Feishu ecosystem, allowing users to build personal work agents without additional development, and its local-first design offers clear advantages in data compliance. For AI application developers, Mem0 provides a more flexible API integration method, suitable for building general-purpose memory layers, but requires manual handling of data integration and conflict resolution. Letta is better suited for dialogue systems requiring long-term memory, such as virtual characters or customer service chatbots, but its architecture has weaker support for non-dialogue data in office scenarios.

In enterprise-level deployment, if the team uses international IM platforms (such as Slack or Teams) or requires cross-platform data aggregation, a traditional RAG solution combined with a vector database may be a more flexible option, although it requires manual implementation of conflict resolution and traceability mechanisms. MyContext's vertical integration capabilities are irreplaceable in office scenarios, but its long-term viability will depend on the speed of ecosystem expansion. It is recommended to conduct a comprehensive evaluation based on data source types, IT operations capabilities, and the need for manual arbitration.

6. Editor's Summary

MyContext is worth noting for its technological innovation: it is the first to concretize the concept of Agent context infrastructure within office scenarios, achieving a complete closed-loop from data ingestion, conflict resolution, to traceable Q&A through a nine-layer architecture. Particularly in conflict information handling, the mechanism of introducing human confirmation and locking in resolution results addresses the long-standing issue of "old facts being erroneously overwritten" in long-term memory systems. This represents a significant differentiating advantage compared to general-purpose memory solutions like Mem0. Additionally, its local-first design inherently supports enterprise data compliance, aligning with the current emphasis on data sovereignty in AI implementation within companies.

In terms of practical value, MyContext effectively lowers the implementation threshold for Agents in office scenarios. Users do not need to write code or configure complex pipelines to have a digital twin that understands their work context. For roles involving dense information (such as project managers and operations personnel), MyContext can significantly reduce the time costs associated with information retrieval and redundant communication. Its open-source nature also allows enterprises to customize and extend it, such as integrating internal knowledge bases or defining custom Agent behaviors.

The primary target audience is individual professionals in the workplace, especially those who use DingTalk or Feishu for daily collaboration. For developers, MyContext provides clear open-source code and architectural documentation, serving as a reference implementation for building memory layers for Agents in vertical domains. Its application scope will further expand after supporting the integration of multi-modal data (such as voice meetings and images) and cross-platform IM systems. Currently, its ecosystem is limited to domestic IM platforms, and international expansion will be a key challenge. However, contributions from the open-source community are likely to accelerate this process.

7. Application Scenarios

• Smart Message Reply: Based on an individual's work history and communication style, MyContext Agent can automatically draft DingTalk/Flybook work messages in the user's tone. The user only needs to review and confirm the message before sending, significantly improving daily communication efficiency. It is especially suitable for handling repetitive messages such as routine reports and progress updates, reducing the time spent on manual typing.

• Work Information Q&A: Users can query their work history using natural language, such as "What was the delivery time confirmed in last week's client meeting?" or "Which project did I collaborate with Zhang San on?" The system returns answers with sources through a hybrid retrieval method, eliminating the need to review chat records or documents, thereby enhancing the efficiency of information retrieval.

• Personal Knowledge Accumulation: Responsibilities, collaboration relationships, and discussion outcomes scattered across IM, documents, and meetings are automatically organized into a continuously updated work profile. Employees can review their historical contributions and decision records anytime, forming a structured personal knowledge base that is useful for annual summaries or project retrospectives.

• Collaboration Relationship Mapping: By visualizing the relationships between people, projects, and topics through a memory graph, the system helps users quickly understand the context of cross-departmental collaboration. For example, a new project manager can clearly see each role's responsibilities and key decision points within the project, accelerating their integration into the team.

• Onboarding for New Hires/Role Transfers: When an employee leaves or changes roles, their work profile is fully traceable. New hires can directly query the previous employee's work records through the Agent, significantly reducing the cost of knowledge transfer and shortening the onboarding period. The traceability of the profile ensures the credibility of the information.

8. FAQ

Q: Where is MyContext's data stored? Is it secure?
A: MyContext uses a local-first architecture, storing all data (including messages, documents, and indexes) on the user's local machine and not uploading it to the cloud. The AI model only functions as a reader of the archives and does not own the data. Users can delete local data at any time, meeting enterprise data security and privacy compliance requirements.

Q: What data sources does MyContext support?
A: MyContext natively supports chat records, documents, and meeting summaries from two major domestic IM platforms: DingTalk and Feishu. In the future, it plans to expand support to additional data sources such as WeCom and Slack. Users can also integrate structured data from other sources via API or file import, but they must handle the format adaptation themselves.

Q: Does using MyContext require an internet connection?
A: An internet connection is needed during the data synchronization phase to retrieve IM and document data. However, once the data is imported locally, subsequent archive building and Q&A reasoning can be performed offline (provided the AI model is locally deployed). It is recommended to maintain an internet connection to receive incremental updates and model upgrades.

Q: How does MyContext ensure the accuracy of Q&A?
A: The system ensures comprehensive information retrieval through three search methods: full-text, semantic, and graph-based. Each answer is accompanied by its original source, allowing users to verify it with a single click. A conflict resolution mechanism further prevents factual contradictions within the archive, enhancing the reliability of the answers. However, the final accuracy still depends on the completeness of the data sources and the model's capabilities.

Q: What is the difference between MyContext and DingTalk's built-in AI assistant?
A: DingTalk's AI assistant primarily provides instant Q&A based on a general large model, lacking long-term memory of the user's work history. MyContext focuses on building a personalized work archive, enabling the AI to understand the user's responsibilities, collaboration, and decision-making context, thereby offering more tailored and traceable responses. The two can be used complementarily.

9. Project Links

1. Product Official Website: https://mycontext.md/
2. GitHub Repository: https://github.com/openTrinity/mycontext