Tencent Hunyuan Open-Sources Hy Memory 2.0: 30B MoE Agent Memory Model, Framework and MemHub
On October 15–16, Tencent Hunyuan released Hy Memory 2.0, a memory solution for AI Agents positioned as a "private memory butler," open-sourcing three components: a memory model, a Python SDK framework, and the MemHub toolkit.
Confirmed
- The core model Hy-Memory-30B-A3B is now on Hugging Face: a 30B-total-parameter MoE activating roughly 3B per inference, built on the hyv3 architecture, supporting bilingual Chinese-English dialogue, open-sourced under Apache-2.0 with an accompanying paper (arXiv)
- It can automatically extract memories from conversations; the framework handles three scenario types: hierarchical long-term cognitive memory for Chat, and facts + reusable memory for Working scenarios (with corresponding MemHub tools)
- The official claim is "A 3B active model beats Opus 5 at memory" — i.e., a 3B-activation model outperforms Claude Opus 5 on memory benchmarks; the claim comes from the model card and official promotion, is repeated in reposts (@victormustar, @AdinaYakup), but has yet to be verified by independent third-party evaluations
Why it matters
- Agent memory is a key bottleneck in current Agent engineering; Tencent enters with a full "model + framework + tools" stack rather than a standalone model, differentiating itself from the community's RAG/external-storage-dominated memory approaches
- The 30B-total / 3B-active MoE design balances capability with inference cost, and the Apache 2.0 license allows free commercial use, lowering the barrier for Agent developers
2026-10-15 ~ 2026-10-16 · 6 related posts
Primary sources
- [source] Tencent open-sources Hy-Memory-30B-A3B under Apache-2.0 — tencent · 2026-10-15
- [source] Tencent Hunyuan open-sources Hy Memory 2.0: model, SDK and MemHub for agent memory — TencentHunyuan · 2026-10-16
- [source] Tencent Hunyuan open-sources Hy Memory, a 30B MoE memory model for AI agents — AdinaYakup · 2026-10-16
- Tencent releases Hy-Memory-30B-A3B, a 3B-active memory-focused model claiming to beat Opus — victormustar · 2026-10-16
2 near-duplicate retellings: AdinaYakup · paf1138