Tencent Hunyuan has open-sourced Hy Memory 2.0 — a production memory stack for LLM agents that includes a Python SDK (hy-memory), a local-first MemHub app, and Hy-Memory-30B-A3B, a specialised long-term-memory writer model. Artefacts landed mid-October 2026 on GitHub, Hugging Face, PyPI (hy-memory 2.0.x) and arXiv. This is version 2.0, not a first launch: Hy Memory 1.0 shipped in May 2026.
The release itself is confirmed. The benchmark rankings below are the authors’ own claims in a preprint that has not been peer reviewed.
What Hy Memory 2.0 is
The project README describes Hy Memory as “a production-grade dual-system cognitive memory engine and Python SDK (hy-memory) for self-evolving LLM agents,” with specialised pipelines for three scenes: Chat, Working and Coding.
Built on that engine, MemHub is “a local-first desktop and CLI product that unifies conversation history and coding traces from user-authorized AI applications (such as Claude Code, Cursor, Codex, OpenCode, and Qwen Code) into a single managed runtime and local data directory.” The pitch is governed cross-application recall from one local data directory, not a cloud memory silo.
The Hugging Face model card and paper position Hy-Memory-30B-A3B as a compact writer that sits on the write path of external memory frameworks — deciding what to retain, how to represent it, and how new evidence updates a persistent bank — rather than as a general chat model. The card says it has ~3B active parameters (the “30B-A3B” naming) and can replace the general-purpose LLM inside Mem0, MemOS and Hy-Memory. License on the model card and README badge: Apache-2.0.
Authors’ benchmark claims (20% subsamples)
The arXiv abstract states:
We present Hy-Memory-30B-A3B, a compact long-term-memory writer trained across Mem0, MemOS, and Hy-Memory.
And:
On fixed 20% benchmark subsamples, Hy-Memory-30B-A3B ranks first on all four evaluated harnesses, including held-out EverMemOS, with a 55.49% macro average versus 54.62% for Claude Opus 5.
Those are author-reported scores on fixed 20% subsamples, with harnesses and the answer/judge stack held fixed while only the writer changes. The paper also reports per-harness means (Mem0 54.02, Hy-Memory 60.55, MemOS 50.51, EverMemOS 56.89) and notes that margins versus Opus 5 on MemOS and Hy-Memory are tiny (0.01–0.02 points) under this deterministic sampled setup — the clearer claimed gap is on held-out EverMemOS (+2.86 points versus Opus 5 in their table).
Do not treat 55.49% as an independent leaderboard result.
How they say they trained the writer
The technical report describes two stages: Memory On-Policy Distillation (OPD), teaching framework-specific write protocols on student-generated banks, then Memory RL with dense per-chunk probes, attributed cross-chunk rewards, over-storage controls and chunk-level GRPO. Training spans Mem0, MemOS and Hy-Memory Chat/Working/Coding; EverMemOS is held out for evaluation. That training story is the authors’; TSN has not reproduced it.
Why it matters
Agent memory is becoming its own product layer: extraction, consolidation, coding traces and cross-app recall. Most of that stack still hangs a general LLM behind a long prompt. A specialised open writer plus an SDK and a local MemHub is a bet that the write policy deserves its own model — and that Chinese labs will keep shipping open agent tooling in the same week as peers (see Alibaba’s OpenGCUA computer-use stack).
What’s still unknown
- Full-benchmark (100%) scores under the same harnesses
- Independent reproduction of the 20% subsample matrix
- How MemHub handles permissions and retention in practice across Claude Code, Cursor and Codex
- Peer review of the training and evaluation claims
What this means
- Tencent Hunyuan open-sourced Hy Memory 2.0 (SDK, MemHub, Hy-Memory-30B-A3B writer); 1.0 was May 2026.
- Model card: ~3B active parameters; Apache-2.0; drop-in writer for Mem0, MemOS and Hy-Memory.
- Authors claim 55.49% macro average vs 54.62% for Claude Opus 5 on fixed 20% benchmark subsamples across four harnesses including held-out EverMemOS — preprint, not peer reviewed.
- MemHub: local-first desktop/CLI unifying authorised app histories (Claude Code, Cursor, Codex, OpenCode, Qwen Code).
Sources
- Fei et al. / Tencent Hunyuan, “Hy-Memory-30B-A3B: Attributed Reinforcement Learning for Long-Term Memory Writing”, arXiv:2610.22196, 15 October 2026
- Tencent-Hunyuan/Hy-Memory on GitHub
- tencent/Hy-Memory-30B-A3B on Hugging Face
- Hy Memory project site
Related stories
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
