Tencent Unveils Team Memory: AI Agents Share Context in Teams

Chinese tech giant Tencent has unveiled Team Memory, a new open-source project that enables teams of AI agents to share context through a unified storage system. The solution is designed to eliminate the problem of context loss in long sessions, where agents must repeatedly reacquaint themselves with users or projects.

A standout feature of Team Memory is its “persona layer,” which creates a stable user representation based on preferences, decisions, and interaction history. According to Tencent, integrating this layer boosts agent accuracy from 48% to 76% after prolonged use, representing a 59% relative improvement.

Unlike traditional approaches where context is transmitted as a large block in each request, Team Memory utilizes four types of reusable assets: Chat Memory (interaction history), Skill (procedures and workflows), LLM-Wiki (structured documentation), and Code-Graph (code and relationship index). Each agent receives only the assets it requires, minimizing redundancy and enhancing security.

However, despite these advantages, experts point to a critical issue: governance. The current version of Team Memory lacks mechanisms to correct inaccurate data or remove outdated facts. This means a single error, once recorded in the shared storage, automatically spreads to all agents interacting with it, potentially causing severe consequences in production environments.

Industry observers are already discussing this challenge on social media. User Blake Murphy noted that “shared context makes data recording the most interesting part but fails to address error correction.” Another commentator, Virgil Maro, emphasized that governance is the most complex aspect of such systems, as it requires decisions on which data should never be recorded at all.