project
Hy-Memory - An Agent memory plugin developed by Tencent Hunyuan.
Hy-Memory is a memory plugin developed by Tencent Hunyuan, specifically designed for long-term collaborative agents such as OpenClaw. Through a 6-layer memory framework × System1/System2 dual-system × evolution chain three-layer architecture, it enables agents to truly...
What is Hy-Memory?
Hy-Memory, developed by Tencent Hunyuan, is a memory plugin designed specifically for long-term collaborative agents such as OpenClaw. Through a 6-layer memory framework × System1/System2 dual-system × evolution chain three-layer architecture, it enables agents to truly remember, remember accurately, remember lightly, and understand you better. It ranks first among similar frameworks in both the authoritative benchmarks LongMemEval and PersonaMem, reducing the number of memories by over 70%, increasing information density by 45%, and achieving a write speed 8 times faster than similar frameworks.
Main functions of Hy-Memory
-
Six-layer memory frameworkThe memory is divided into six layers: original traces, atomic facts, identity profiles, conversation summaries, mental models, and forward-looking intentions, with each layer having independent responsibilities and retrieval weights.
-
System1/System2 dual systemSystem1 performs millisecond-level real-time processing of L1-L4 memory writes, while System2 asynchronously accumulates L5-L6 higher-level cognitions in the background.
-
Evolutionary chain mechanismThe supersedes pointer links memories into a causal chain, and hitting any node automatically unfolds the complete attitude evolution.
-
Memory merge deduplicationSimilar facts are automatically merged, and conflicting preferences are automatically refreshed to avoid the coexistence of old and new data, thus preventing the formation of a noisy database.
-
Cross-Session Continuous MemorySupports resuming tasks that were closed yesterday, allowing the agent to seamlessly continue the context of long-running tasks.
Hy-Memory Technology Principles
-
Tiered storage architectureThe process consists of L1 original dialogue, L2 atomic facts, L3 identity profile, L4 conversation summary, L5 mental model, and L6 prospective intent, with different processing strategies for each layer.
-
System1 shortcutWhen a user sends a message, the system writes the original traces, extracts facts, updates the profile, and compresses the conversation summary in real time, achieving a closed loop in milliseconds.
-
System2 Slow PathIt runs asynchronously in the background, abstracting mental models from behavior and predicting forward intentions, accumulating data in seconds to minutes.
-
Evolutionary chain pointer structureWhen a new memory is written, the supersedes pointer points to the old memory, forming a bidirectional traversable causal chain.
-
Attention gate mechanismIn System1, attention criteria are set to determine which information is worth processing and writing into deep memory.
How to use Hy-Memory
- Access methodWith OpenClaw, you can connect to Hy-Memory with one click, upgrading your native memory capabilities to professional-grade long-term memory.
- Lite modeIt only performs writing and retrieval, has zero LLM cost, and offers the fastest access speed, making it suitable for those who want to remember it immediately but do not need to understand it yet.
- Pro ModeMemAgent performs extraction, summarization, and reflection synchronously, but without a background worker, making it suitable for those who want to organize their memories.
- Ultra ModeIt fully runs the System1 + System2 kernel, and continuously replays and summarizes asynchronous slow paths, making it suitable for users who want their systems to become more and more like user-friendly systems over time.
Hy-Memory's core advantages
- Higher memory densityThe density of a single message reaches 130.5 tokens/message, which is 2.5 times that of mem0 and 1.5 times that of Graphiti.
- Fewer memoriesIt averages 82.3 messages per user, only about 1/4 of mem0 and Graphiti, thus solving the problem of memory fragmentation.
- Faster write speedIt achieves 12.3 seconds per k tokens, which is 1/8 of Graphiti's speed, without slowing down the mainnet response.
- Leading in evaluation resultsLongMemEval 85.20% and PersonaMem 76.91%, both ranking first among similar frameworks.
- Complete causal evolutionThe evolutionary chain preserves the complete attitude evolution path, avoiding the problems of the coverage faction only recording the latest information or the accumulation faction recalling incomplete information.
Hy-Memory project address
- Project official websitehttps://memory.hunyuan.tencent.com/
Hy-Memory's Competitive Product Comparison
| Dimension | Hy-Memory | mem0 | Graphiti |
|---|---|---|---|
| LongMemEval Total Score | 85.20% | 47.00% | 68.32% |
| PersonaMem Total Score | 76.91% | 65.82% | 64.86% |
| Write time (s/k tokens) | 12.3 | 15.6 | 97.8 |
| Number of records/users | 82.3 | 309.8 | 362.3 |
| Density of a single message (token/message) | 130.5 | 52 | 89.2 |
| Memory architecture | 6-layer hierarchical structure + evolutionary chain | Single-layer vector | Graph structure |
| Dual-system processing | System1 + System2 | none | none |
Applications of Hy-Memory
-
Long-term project collaborationWhen following up on complex projects over several weeks, the agent can remember the reasons for each decision and the solutions to avoid them.
-
Personal knowledge managementIt accumulates users' work habits and decision-making mental models, and the more it is used, the better it understands user preferences.
-
Fitness/Health PlanningRecord the evolution of attitudes towards training methods to avoid recommending solutions that users have already encountered problems with.
-
Creative assistanceTrack changes in creators' attitudes toward distribution channels and provide recommendations that align with their values.
-
Learning tutoringRemember students' learning progress, knowledge gaps, and comprehension methods, and provide continuous guidance.