Tencent's Team Memory: Shared AI Agent Memory
Team-wide shared AI memory increases the risk of spreading incorrect information.
1The Reliability Issue of AI Agents
A June 2024 survey revealed that 57% of enterprises attributed incorrect AI agent answers to missing or inconsistent context. This raises questions about the reliability of AI agents' independent decision-making capabilities. Current solutions focus on a single agent remembering more in one session. However, enabling a team to share the same context remains a challenge.
2Tencent's Solution: Team Memory
Tencent developed Team Memory to address this issue. This project emerged from six months of work aimed at preventing context loss in long sessions. The system includes a 'persona layer' that provides a stable definition of who the user is and how they work. Adding this layer increased agent accuracy from 48% to 76%.
3Shared Memory System
Team Memory allows all agents in a team to read from the same memory hub. The system registers four types of reusable assets and equips each agent only with what it needs. These assets include chat memory, skills, LLM-Wiki, and code graph. This approach prevents every agent from accessing everything.
4Access Control Layer
Team Memory determines who can read what through an access control layer. There are four visibility levels: private, team, restricted, and agent. New assets are set to private by default, making sharing a deliberate action. This model answers the question of who is allowed to read a specific memory asset.
5The Problem of Incorrect Memory
However, the question of what happens when a memory asset turns out to be wrong remains unanswered. Tencent's documentation explains ownership, versioning, and status tracking for each asset but does not define a process for correcting or expiring incorrect facts. This gap was quickly noticed by practitioners.
6Risk of Spreading Incorrect Information
Shared memory carries the risk of incorrect information spreading to all team members' agents once written. This issue becomes more serious than a single user having to repeat an explanation. Experts express concerns about how correction and expiration processes will be managed in such cases.
7Conflicting Memories
The question of what happens when two agents have conflicting memories about the same thing also remains unanswered. In shared memory mode, it is unclear whose memory will prevail. This situation could cause shared memory to drift quickly compared to the slow drift of a single agent's memory.
8Future Challenges
Tencent's Team Memory has the potential to turn AI agents into a real team. However, governance, especially when facts conflict, will be the most challenging part. Experts point out structural risks in shared multi-agent memory systems, such as governance fragmentation and silent quality degradation without feedback loops.
Original article: VentureBeat — Tencent'in Takım Hafızası: AI Ajanları İçin Paylaşımlı Bellek