Tencent’s Team Memory Unveils AI Agent Memory Sharing for Teams, Lacks Oversight for Inaccuracies

A recent survey conducted by VB Pulse in June revealed that 57% of companies identified instances where AI agents provided incorrect responses due to a lack of or inconsistent context. This highlights the increasing importance of context in determining the reliability of AI agents’ autonomous actions.

Current solutions have primarily addressed a more limited scope, focusing on individual agents retaining context during a single session. However, a significant challenge remains: enabling a group of agents to utilize shared context simultaneously. When an agent’s context is collective, an inaccurate piece of information can affect not just one individual, but the entire team.

Tencent has introduced a solution to bridge this gap with its Agent Memory, an open-source initiative developed over six months to tackle the issue of agents losing context in extended sessions. This system incorporates a persona layer, which creates a stable representation of a user built from numerous interactions instead of being recreated each time. In Tencent’s evaluations, the accuracy of agents applying this persona over time improved from 48% to 76% following the integration of the persona layer. Recently, Tencent expanded this initiative with the beta launch of Team Memory, facilitating this approach for entire teams rather than individual agents. The project notably reached the top of GitHub’s TypeScript trending list this week.

With Team Memory, agents can reference a collective memory hub instead of maintaining isolated contexts. An access control layer manages who can access specific information.

Team Memory’s Functionality

The central concept of Team Memory revolves around a shared hub instead of a single prompt. Rather than replicating extensive context for each agent, Team Memory organizes four types of reusable elements, equipping agents with only what they require:

• Chat Memory: Preserves user preferences, facts, choices, and interaction history, processed through four stages from raw conversation to a stable, long-term persona, ensuring an agent does not need to be reintroduced to a user it has previously engaged with.

• Skill: Compiles methods derived from completed tasks, which are versioned and reviewed before becoming accessible.

• LLM-Wiki: Converts documents and specifications into structured, interconnected pages.

• Code-Graph: Catalogs a codebase’s components and their interrelationships, enabling agents to assess the implications of potential changes beforehand.

Tencent emphasizes the distinction: “RAG identifies ‘what can be accessed?’ while Team Memory clarifies ‘who can utilize it, which version is authoritative, and which agent should have it.'”

This leads to what Tencent defines as an “Agent Loadout”: for instance, a Scout agent can be given market research resources, while a Builder agent receives necessary documentation, ensuring that agents only access relevant information.

Access to assets is managed through four visibility tiers:

• Private: Only the asset’s owner can view it.

• Team: Accessible to all team members.

• Restricted: Access controlled by user, role, or agent-specific permissions.

• Agent: Limited to a specific agent in the team.

New assets default to private, requiring intentional sharing instead of automatic distribution.

Addressing Memory Errors

This access model effectively addresses who can view a memory asset, but does not clarify what happens when an asset is proven to be incorrect. While Tencent provides documentation covering ownership, versioning, and status tracking, it lacks guidelines for correcting or invalidating previously accessed facts, or resolving conflicts between agents that remember differently.

Practitioners raised concerns shortly after the initial announcement, highlighting the complexities introduced by shared memory: a single error can propagate to every team member’s agent. This shift from individual to collective memory may amplify the need for stringent governance processes to manage corrections.

Some questions raised included how to navigate situations where two agents have conflicting memories about the same information. The potential for misinformation to spread rapidly within a team is a significant concern, as once faulty information is recorded, it could impact all agents relying on that shared pool.

Despite the criticisms, some noted a positive transition: transforming memory into a shared resource fosters collaboration among agents, rather than isolating them. The challenges associated with governance, particularly when faced with contradictory facts, were recognized as crucial points to explore further.

These issues aren’t unique to Tencent’s implementation. A recent independent paper on multi-agent memory architecture raised alarms about governance fragmentation and the risk of quality deterioration due to shared memory systems. The discrepancies noted also align with what commenters mentioned: errors in a single-agent memory system primarily burden one user, whereas in a shared system, they can affect the entire team.

Comparing Team Memory

In 2026, advancements in AI agent memory have primarily been focused on enhancing individual agents’ recall capabilities. Various initiatives like LangChain’s LangMem SDK and Google’s Always On Memory Agent have been developed with these goals in mind. A different avenue of development has been aimed at providing agents with a shared understanding of business data. Reports suggest that only 25% of companies have this kind of context layer implemented in production.

The closest comparison to Team Memory may be Asana, which has developed a similar framework for its AI tools, preventing agents from needing to be constantly re-briefed on context that other agents already possess. Both platforms are contending with analogous access control challenges, ensuring that one agent’s memory does not inadvertently influence unrelated projects. However, Tencent’s initiative stands out due to its open-source nature, allowing cross-framework portability.

For organizations assessing developments in this area, the benefits are clear: agents can avoid having to relearn collective information. However, the associated risks of communal memory, such as the potential widespread impact of a single inaccuracy, emphasize the necessity for refined corrective mechanisms and governance practices.