Tencent Hunyuan Open-Sources Hy-Memory Agent Memory Stack

When agents forget, the bottleneck is often the write path, not the retriever: they skip what matters, append instead of revise, and leave contradictory facts piled up. Tencent Hunyuan’s Hy-Memory 2.0 open-sources three layers—model, framework, and app—aimed squarely at long-term memory.

The model is Hy-Memory-30B-A3B, a roughly 30B-total MoE memory writer with about 3B active parameters. It is meant to replace the general-purpose LLM that usually sits behind Mem0, MemOS, or Hy-Memory and decides what to keep, how to phrase it, and how new evidence should update the bank. Training combines Memory On-Policy Distillation with Attributed Memory Reinforcement Learning. On the reported fixed subsamples, the model’s four-harness macro average is about 55.49, ahead of Claude Opus 5 at about 54.62, and it still leads on held-out EverMemOS. Weights are on Hugging Face as tencent/Hy-Memory-30B-A3B.

The Hy-Memory SDK frames three memory scenes: chat memory that lifts scattered facts into event timelines; working memory that tracks where a task stands and when experience applies; and coding memory that records why a change was made. On Hunyuan’s published open-framework comparisons, Hy-Memory claims first place across chat, working, and coding benchmarks while keeping storage relatively tight. MemHub, the app layer, wires that memory into tools such as Cursor, Codex, and OpenCode—import history, sync continuously, recall through plugins—so preferences are not trapped inside one client.

One timeline note: Hunyuan’s official site dates this open release around October 16, 2026. It should be read from those official pages, not as a same-day X first post from a community relay. What matters for builders is that the doors are open—Tencent-Hunyuan/Hy-Memory on GitHub, HF weights, the technical report, and MemHub. Long-horizon agents eventually compete less on another chatty base model and more on whether they rewrite their own memory honestly. Making “write” a specialized small model is the engineering bet worth watching.