Tencent's Team Memory shares AI agent memory across a team — with no governance yet for when it's wrong
A VB Pulse survey this June found that 57% of enterprises had traced a confidently wrong agent answer back to missing or inconsistent context — the latest sign of how central context has become to whether AI agents can be trusted to act on their own. Most of the fixes so far have solved a narrower version of that problem: one agent remembering more, in one session. What's been missing is a way…
A recent survey indicates that 57% of enterprises have identified confidently incorrect responses from AI agents as being due to missing or inconsistent context. This highlights the growing importance of context in determining whether AI agents can autonomously function. While current solutions primarily address agents retaining more context within a single session, a new issue has emerged: when context is shared across a whole team of agents.
Tencent has responded to this challenge with Agent Memory, an open-source project developed after six months of work aimed at resolving the problem of agents losing context during extended sessions. A key component of this system is a persona layer, which establishes a stable, distilled representation of a user's identity and work habits through multiple interactions rather than reconstructing it each time.
This enhancement has been shown to improve an agent's accuracy from 48% to 76% when applying this persona over extended use.
Building upon this foundation, Tencent has now launched Team Memory, a beta version that extends the shared context approach to an entire team of agents. Instead of duplicating context blocks within each agent's window, Team Memory utilizes a shared memory hub. This hub registers four types of reusable assets: Chat Memory, Skill, LLM-Wiki, and Code-Graph. These assets are then assigned to agents based on their specific roles and needs through an access control layer that determines visibility.
The core concept behind Team Memory is to maintain a shared hub rather than a shared prompt. Raw conversation data is distilled into Chat Memory, capturing conversation preferences, decisions, and interaction history through four layers, including a stable long-term persona. The Skill component captures procedures from completed work, with versioning and review before sharing.
LLM-Wiki converts documents and specifications into structured, linked pages, while Code-Graph indexes a codebase's symbols, files, and call relationships to help agents understand the impact of changes before making them.
While this approach offers the advantage of team-wide context sharing, concerns have been raised regarding governance when information within these shared memories becomes incorrect. Tencent's documentation outlines ownership, versioning, and status tracking for each asset, but it does not provide a clear process for correcting or expiring factually erroneous information once it has been read and reused by multiple agents on a team.
This gap has been highlighted as a critical issue, with practitioners questioning how the governance layer will address memory correction and expiry, especially when multiple agents' memories contradict each other.
Written by urgent.news from VentureBeat's reporting — not their text. Machine-written — it may contain errors, so check the original before relying on it.
