Tencent Unveils Hy Memory 2.0: Long-Term Memory for AI Agents

On October 15, 2026, Tencent Hunyuan released Hy Memory 2.0 across three layers: the Python SDK hy-memory==2.0.1 on PyPI, a monorepo featuring the MemHub application on GitHub, and the MoE model Hy-Memory-30B-A3B on Hugging Face (Apache-2.0) — for recording the long-term memory of AI agents.

Compiled by Sergey KostenchukPublished 2026-10-16Updated 2026-10-16

2026-10-16 Coding Anthropic

🧠 Tencent Unveils Hy Memory 2.0: Long-Term Memory for AI Agents

Tencent Hunyuan has released Hy Memory 2.0: the hy-memory==2.0.1 SDK on PyPI, a monorepo with MemHub on GitHub, and the Hy-Memory-30B-A3B model (Hugging Face, Apache-2.0). The 30B MoE with 3B active parameters is trained to record long-term memory; it claims first place on four memory frameworks (55.49% vs. 54.62% for Claude Opus 5).

🌍 Long-term agent memory has become a separate trainable layer instead of a general-purpose LLM behind a prompt; Apache-2.0 and the compact model simplify integration with Mem0 and MemOS.

👤 You can install hy-memory==2.0.1, download the weights, and run MemHub (CLI, Mac): context is described once and works across 11 tools — Claude Code, Cursor, Codex, OpenCode, Qwen Code. With 3B active parameters, local deployment is realistic; the synchronization mode for personal data is not disclosed.

Source 1: https://github.com/Tencent-Hunyuan/Hy-Memory Source 2: https://huggingface.co/tencent/Hy-Memory-30B-A3B

Sources