Tencent Hunyuan Unveils Hy-Memory: A 'Second Brain' for Long-Term AI Agents

IT Home reported on May 28, 2026, that Tencent Hunyuan has officially launched Hy-Memory. Described as a plugin for the Openclaw agent, Hy-Memory aims to function as the agent's "second brain," specifically engineered to handle the complexities of long-term collaboration.

Solving the 'Three-Week Trajectory' Problem

The release addresses a common phenomenon in the AI interaction lifecycle known as the "three-week trajectory":

• Week 1 (Honeymoon): Users are enthusiastic, sharing complex project details and future goals with the agent.
• Week 2 (Confusion): Users begin to struggle, having to re-explain context or reasons for previous decisions as the agent's memory drifts.
• Week 3 (Downgrade): Users stop asking deep questions, reducing the agent from a "thinking partner" to a simple "query tool." Hy-Memory aims to eliminate these second and third weeks, ensuring the agent retains context and understanding indefinitely.

Core Architecture and Standards

To qualify for long-term collaboration tasks, Hy-Memory was built on three strict standards:

1. No Loss of History: Crucial judgments and causal reasoning cannot be discarded.
2. Evolution: The system must evolve alongside the user, not just overwrite old data with new data.
3. Performance: Memory retrieval must be instantaneous and not bottleneck the main user workflow.

1. The 6-Layer Memory Framework

Traditional memory systems often mix distinct types of information into a single vector pool, making retrieval inefficient. Hy-Memory categorizes memories into six layers, each with a specific responsibility:

• L1-L4 (System 1): Raw traces, facts, and session summaries written in real-time.
• L5-L6 (System 2): Abstracted personas, intent models, and knowledge networks built over time.

This separation ensures that when a user asks a question, the agent retrieves the most relevant layer (e.g., a simple location query uses L2, while a decision-making query uses L5), keeping the prompt clean and effective.

2. System1/System2 Dual-System Design

Inspired by cognitive science, Hy-Memory splits memory processing into two channels:

• System 1 (White Shift): Handles immediate input, writing raw traces and updating profiles instantly so users don't experience lag.
• System 2 (Night Shift): Runs in the background to perform deep cognitive tasks like building knowledge networks and predicting user intent.

This allows the agent to remember user details immediately while continuously improving its understanding in the background.

3. The Evolutionary Chain

The most innovative feature is the "Evolutionary Chain." Instead of overwriting past data or keeping a chaotic pile of everything, Hy-Memory uses "supersedes" pointers. When a user's preference changes (e.g., shifting from running to weightlifting due to injury), the new memory points to the old one.

This creates a linked chain of history. When the agent retrieves information, it retrieves the entire timeline of the user's evolution, allowing for nuanced, context-aware responses rather than shallow, repetitive ones.

Performance and Benchmarks

Hy-Memory has demonstrated superior performance in both efficiency and capability:

• Memory Density: 45% higher information density compared to existing frameworks.
• Efficiency: Memory count reduced by 70%+; token consumption lowered by 35%.
• Speed: Update speed increased by 20%.

In authoritative tests like LongMemEval and PersonaMem, Hy-Memory scored 85.2 and defeated all competitors, particularly excelling in preference evolution, temporal reasoning, and knowledge updates.

By reducing noise in the prompt and preserving the full context of user evolution, Hy-Memory aims to transform the AI agent from a temporary assistant into a permanent, evolving partner.

Conclusion

With Hy-Memory, Tencent Hunyuan is tackling the fundamental challenge of long-term AI memory. By preserving the "why" and "how" of past decisions alongside current preferences, the system ensures that agents like Openclaw can grow with their users, maintaining the utility of the first week of interaction well into the future.