Tencent has released Hy-Memory, a memory plugin designed for long-term collaborative agents like OpenClaw. It tackles the classic problems of memory fragmentation, attention dilution, and hallucination that plague long-running AI systems. The goal: agents that actually remember, recall correctly, stay lightweight, and understand you better across days and sessions.
On the technical side, Hy-Memory ditches the single vector table in favor of a six-layer memory framework that splits memories into facts, profiles, mental models, and forward-looking intentions. To balance speed and depth, it mimics the human brain's dual-system: System 1 handles real-time millisecond retrieval of immediate facts and conversation summaries (layers L1–L4), while System 2 runs in the background, taking seconds to minutes to distill user mental models and knowledge networks (L5–L6). The result is zero latency for the main conversation.
For constantly changing preferences and facts, Hy-Memory introduces an "evolution chain" that links old and new memories via supersedes pointers. When the agent searches, it automatically unfolds the entire chain. This ensures the agent gets the latest conclusion while preserving the full causal path of user decisions — say, from steady cardio to HIIT, a knee injury, and finally a mix of training styles. No more recommending ideas the user already rejected.
On the long-term memory benchmark LongMemEval, Hy-Memory scored 85.2, leading in evolution metrics like preference tracking, temporal reasoning, and knowledge updates. It also topped the long-term real conversation benchmark PersonaMem. Performance-wise, Hy-Memory writes 8x faster than Graphiti, uses about a third of the memory entries of mem0 and a quarter of Graphiti's, and packs 45% more information density per memory. In long-context processing, it cuts token consumption by 35% and speeds up memory updates by 20%.
Hy-Memory ships in Lite, Pro, and Ultra tiers, with single-command integration. It defaults to Chrome's local embedded vector library, no need for Qdrant or Docker. The same SDK supports seamless upgrades.