A June VB Pulse survey found 57% of enterprises traced a confidently wrong agent answer to missing or inconsistent context. Tencent's response is Team Memory, an open-source beta that extends its existing Agent Memory project into a shared hub for entire agent teams. The underlying Agent Memory system already showed results: adding a persona layer lifted long-session accuracy from 48% to 76% on Tencent's own benchmark, a 59% relative gain. The repo hit number one on GitHub's TypeScript trending list this week.
Team Memory registers four asset types, Chat Memory, Skill, LLM-Wiki, and Code-Graph, and equips each agent with only what it needs through a four-tier access control system: private, team, restricted, and agent-level. New assets default to private. That architecture separates Team Memory from RAG: Tencent's own documentation states RAG answers 'what can be found,' while Team Memory also answers 'who can use it, which version is valid, and which agent should receive it.' A Scout agent gets market research assets. A Builder agent gets the code graph. Nothing is dumped into every agent's window at once.
The unresolved problem is what happens after a wrong fact enters the shared hub. Tencent's documentation covers ownership, versioning, and status tracking, but describes no correction or expiry process for a fact already read and reused across a team. Practitioners flagged this within hours of launch. Austin Green noted on X that single-agent memory drifts slowly while shared memory drifts fast, because one stale write propagates to agents that never saw the session that produced it. A March 2026 paper, 'Governed Memory: A Production Architecture for Multi-Agent Workflows,' identifies exactly this structural risk independently of any vendor. The architecture is worth reading in full for how it frames the write path, not retrieval, as the hard problem in multi-agent memory.
[READ ORIGINAL →]