Tencent Hunyuan Launches Hy-Memory Agent Memory Plug-in

Tencent Huyuan officially launched Hy-Memory on May 28th, positioning it as a memory plugin for long-term collaborative Agents such as Openclaw. The official summary of its core architecture includes a 6-layer memory framework, System1/System2 dual system, and evolutionary chain triple-layer design, with the goal of allowing Agents to "remember accurately, remember correctly, remember lightly" during long-term use.

The core logic of Hy-Memory is to store memories in layers, rather than mixing all information into the same vector library. The 6-layer framework corresponds to raw dialogue, facts, user portraits, session summaries, mental models, and knowledge networks, with different types of memories taking different retrieval paths to avoid irrelevant content diluting the model's attention.

The dual system design addresses the contradiction between speed and depth. System1 completes real-time writing in an instant when a user sends a message, while System2 processes deeper cognitive abstractions in the background asynchronously, such as extracting users' decision-making habits and long-term preferences. The two channels run in parallel without blocking the main dialogue thread.

The evolutionary chain is a differentiated capability that the officials emphasize. The system strings historical memories under the same topic into a chain using pointers, with new memories not directly overwriting old ones, but preserving the complete evolutionary trajectory. Officials give an example, stating that this can prevent Agents from recommending similar high-intensity exercises to users who have been injured by HIIT training in the past.

In the LongMemEval test set, Hy-Memory scored 85.2, leading similar products by 21.11, 9.63, and 21.37 percentage points in preference tracking, temporal reasoning, and knowledge updating, respectively. In terms of performance, the write speed is 8 times that of Graphiti, the number of memory entries is only one-third of mem0, but the density of individual memory information is 3 to 4 times higher.

Hy-Memory is currently provided in the form of a plugin, supporting local embedded deployment, with the default use of the Chroma vector library, no external services or Docker environment required, and officials say the installation process takes about 5 minutes. The product is divided into Lite, Pro, and Ultra three levels, enabling different levels of memory capabilities as needed.