Alibaba's Qwen Office open-sources MyContext to turn DingTalk chats and documents into agent context

Alibaba is turning context into agent infrastructure. On August 18, QbitAI reported that Alibaba's Qwen Office team open-sourced MyContext, a context infrastructure project that continuously processes DingTalk chats, enterprise documents, meeting records and collaboration data into work context that agents can directly consume. The project has already passed 1,000 stars on GitHub just one week after going open source.

The release targets a painful gap in real-world agent adoption. While model reasoning, coding and tool-calling keep advancing, and harnesses, plugins and execution frameworks fill in task orchestration and state management, agents still know almost nothing about actual business context: what was discussed with colleagues in DingTalk, which decisions are recorded in documents, and what standards a company follows often sit entirely outside the agent's reach.

MyContext's answer is direct: it adds a data processing layer for agents. With user authorization, IM conversations, documents, meetings and collaboration records are continuously collected and organized into a dynamically updated work profile. Which projects a user owns, who they collaborate with regularly, and which discussions have concluded are systematically retained and fed into downstream task execution, so users no longer have to reintroduce themselves every time they switch agents.

For the messy realities of enterprise data — late arrivals, constant changes and contradictory facts — MyContext applies engineering solutions. Each piece of raw information is bound to a stable source identifier so idempotency guarantees late-arriving data is not lost; idle gaps in conversation define session boundaries so context segmentation follows real interaction rhythm; and a sliding time window continuously aggregates evidence, so facts that recur across different discussions generate new confidence signals.

To handle factual conflicts, MyContext uses a three-state merge mechanism that treats conflict as a business signal worth preserving: consistent information boosts confidence, supplementary information merges into existing conclusions, and genuine conflicts keep multiple facts with lowered confidence, exposed explicitly to the user. Conclusions manually confirmed by users get higher priority and cannot be overwritten by the model later.

The project also avoids turning context into a black-box memory. Every conclusion keeps a traceable evidence chain back to the original chat, document or meeting, and everything the agent can see and call is constrained by user and organizational permissions. The team also tackled the cost problem of context engineering, so a new batch of messages does not force a full recomputation of all history.

Overseas vendors have already explored this territory: Palantir unifies business objects with Ontology, Glean emphasizes its Enterprise Graph, and Microsoft connects enterprise data and permissions through Microsoft Graph and Copilot Connector. MyContext differentiates itself by going deeper into the engineering problem — turning heterogeneous, strongly time-ordered, continuously changing and even conflicting raw business data into stable, agent-consumable context. QbitAI cites a Confluent 2026 survey finding that 66% of enterprises say data infrastructure and quality are slowing agentic AI adoption, and 80% have made using their own data to drive AI a business priority.

The bigger signal is that enterprise AI adoption is no longer blocked by whether a stronger model exists, but by whether there is backend infrastructure that can continuously govern and process scattered business data into usable context. Passing 1,000 GitHub stars in a week shows broad developer agreement on the pain point of agents starving for real business context. The next things to watch are whether MyContext builds an ecosystem, gets integrated into mainstream harnesses and agent frameworks, and how deeply Alibaba connects it with its Qwen product line.