Tencent Hunyuan open-sources Hy Memory 2.0: a ~3B-active memory-writer model (Hy-Memory-30B-A3B) that edges Claude Opus 5 in its own memory benchmarks, plus an SDK and MemHub cross-tool memory app

Why it matters

【Key points】

On October 16 Tencent Hunyuan released Hy Memory 2.0 and open-sourced three layers under Apache 2.0: Hy-Memory-30B-A3B, a dedicated memory-writer model (30B total, ~3B active per token); the hy-memory 2.0 Python SDK; and MemHub, a local-first desktop/CLI app. Weights (plus FP8) went up on Hugging Face on Oct 15–16 with an arXiv technical report.

The model does not answer questions. It handles the write path of an external memory system: what to keep, how precisely, and how new evidence updates the bank. It drops into Mem0, MemOS and Hy Memory in place of a general LLM. Training is on-policy distillation followed by chunk-level attributed RL.

In Tencent's own tests across four memory frameworks (EverMemOS held out) and six metrics, it averages 55.49 vs Claude Opus 5 at 54.62, DeepSeek V4.1 Flash 53.64 and Qwen3.6-35B-A3B 50.37. MemHub merges history from Claude Code, Cursor, Codex, OpenCode, Qwen Code and others into one local memory. The repo has ~800 stars; community GGUF/MLX/AWQ quants exist.

【Verifiable】

Model card: https://huggingface.co/tencent/Hy-Memory-30B-A3B

Code + MemHub: https://github.com/Tencent-Hunyuan/Hy-Memory

Tech report: https://arxiv.org/abs/2610.22196

【Our read】

1. Self-reported, on 20% subsamples, and under 1 point ahead of Opus 5 — effectively a tie until replicated.

2. The idea matters more: hand the costly, error-prone memory write step to a ~3B-active specialist instead of a flagship model, which cuts cost for long-running agents.

3. MemHub's cross-tool memory competes with vendors' built-in memory; it keeps data local, but users own the consent and privacy boundary.