Tencent Hunyuan Unveils Hy-Memory: 45% More Memory, 35% Lower Token Cost

Long-term collaborative AI agents often experience a frustrating "three-week trajectory": the first week impresses, the second week brings unease as memory lapses and drifting understanding become apparent, and by the third week, they degrade into basic instant query tools. To address the challenges of fragmented memory and chaotic timelines in extended collaboration, Tencent Hunyuan today (May 28, 2026) launched Hy-Memory, a smart memory plugin built for long-term collaborative agents like Openclaw — Hy-Memory.

As the "second brain" of an agent, Hy-Memory outperforms mainstream frameworks on the authoritative public benchmarks LongMemEval and PersonaMem, thanks to three core technologies. Evaluation data shows that Hy-Memory reduces memory volume by over 70% while increasing per-memory information density by more than 45%; for long-context processing, token consumption drops by 35%, and memory update speed improves by 20%.

Three Core Technologies Reshaping the AI Memory Mechanism

Layer One: The Six-Level Memory Framework (Precise Placement)

Hy-Memory avoids dumping all information into the vector database at once. Instead, it decomposes memory into six levels: L1 raw traces, L2 atomic facts, L3 identity profiles, L4 session summaries, L5 mental models, and L6 prospective intentions. It precisely invokes the relevant level based on user queries, streamlining prompts and preventing irrelevant text from diluting the model's attention.

Layer Two: The System1/System2 Dual Systems (Balancing Speed and Depth)

Memory processing is split into two independent channels: System1 (day shift mode) extracts facts and updates summaries (L1-L4) in real time, within one second of the user pressing enter, so the information is ready for the next conversation; System2 (night shift mode) runs asynchronously in the background, taking seconds to minutes to deeply consolidate the user's mental models and knowledge networks (L5-L6), making the agent smarter over time without blocking the main process.

Layer Three: The Evolutionary Chain Mechanism (Preserving Causal Traces)

This is Hy-Memory's secret weapon. When users change their views or habits, traditional systems either "overwrite the old with the new," losing historical experience, or "pile up disorderly," causing incorrect recall. Hy-Memory links new and old memories into an "evolutionary chain" via the supersedes pointer. Once any node in the chain is triggered, the entire evolutionary path unfolds automatically. The agent not only remembers the user's latest conclusion but also revisits the full evolution of their attitude, thereby avoiding past "pitfalls."

Extreme Performance and 5-Minute Rapid Deployment

For performance, Hy-Memory achieves ultra-fast write speeds comparable to mem0 (8 times faster than Graphiti), while using only one-third the number of memories. It uses local embedded storage by default, and no Docker deployment, external services, or Qdrant database are required, with data automatically persisted locally.

To adapt to different hardware and scenario needs, Hy-Memory offers Lite (Lightweight), Pro (Professional), and Ultra (Ultimate) configurations. These three tiers share the same SDK, and switching between them is seamless by simply changing switches. The project has officially launched, and users can find more guides through its official website or user documentation