Tencent Unveils Team Memory, a Shared Memory System for AI Agents, in Beta

Chinese IT giant Tencent has released a beta version of Team Memory, an open-source system that enables multiple AI agents to share common memories and context. Rather than individual agents holding information separately, the entire team can reference a shared "memory hub." The system topped GitHub's TypeScript trends this week.

The project originated from a fundamental challenge: AI agents lose contextual information during extended conversation sessions. Tencent's team spent approximately six months addressing this problem, which led to the creation of Agent Memory, the predecessor to Team Memory. The core component, the "persona layer," distills user characteristics and work patterns from accumulated conversations into stable profiles. By adding this feature, Tencent's proprietary benchmark improved accuracy from 48% to 76% (a relative improvement of approximately 59%). Team Memory extends this approach to the team level.

Team Memory manages four types of shared assets. "Chat Memory" condenses user preferences and decisions from conversation history into long-term personas. "Skills" extract procedures from completed tasks and are shared only after version control and review. "LLM-Wiki" converts documents and specifications into structured pages, while "Code Graph" indexes relationships within the codebase to preemptively identify the scope of changes. Rather than providing all assets to each agent, the design assigns only role-specific asset sets ("agent rollout") to each agent.

Tencent's documentation explains the difference from traditional retrieval-augmented generation (RAG) this way: "RAG answers 'what can be found.' Team Memory further answers 'who can use it,' 'which version is valid,' and 'which agent should receive it." Access control layers manage read permissions, allowing control over the information agents can reference.

This initiative gains importance in light of reliability challenges in AI agents. A survey by VentureBeat in June this year found that 57% of enterprises have experienced instances where they traced incorrect agent responses back to missing or inconsistent contextual information. Simply increasing a single agent's memory capacity proves insufficient for addressing issues in team environments where multiple agents operate collaboratively.

However, a significant gap remains: governance mechanisms for handling misinformation within shared memory—specifically, who corrects it, when, and how verification occurs—have not yet been established. Unlike errors confined to a single agent, misinformation shared across an entire team could have far broader consequences. As Team Memory is released as open source, how the community and enterprises address these governance considerations during implementation will be a key point of attention moving forward.