Tencent Hunyuan open-sources Hy Memory 2.0 and a 30B memory model

Tencent Hunyuan’s Hy Memory 2.0 page, dated 2026.10.16 on the page, says the release is open at three layers: a memory model, a framework, and an app. The GitHub README calls the October release Hy Memory 2.0, package hy-memory==2.0.2, and says the monorepo and MemHub are open-sourced together. The LICENSE file is the Apache License 2.0, with a 2026 Tencent copyright line above it. The file says Hy-Memory adds no restrictions beyond that license, and that the name covers code, parameters, and weights Tencent makes public under it.

PyPI’s 2.0.2 record lists license Apache-2.0, Python 3.9 or newer, and upload time 2026-10-16T16:01:45Z. The author field is alvinfei. The name is older: 1.0.0 uploaded on 21 May 2026, and 2.0.1 on 15 October 2026.

Three scenes in the SDK

The README describes three scenes. Chat is hierarchical long-term memory. Working stores facts and reusable experience. Coding tracks a version-evolution chain. A write with a tool-call trace can be stored as working or coding memory in pro or ultra mode. The README says lite ignores the scene and stores the text as raw memory. The install line is pip install hy-memory==2.0.2. The quick start says any OpenAI-compatible LLM and embedding endpoint can be set with environment variables. Sample values in the README point at a DeepSeek chat model and an Alibaba embedding URL.

The Hugging Face card names Hy-Memory-30B-A3B, license Apache 2.0, about 3 billion active parameters. It says the model replaces the general-purpose writer inside Mem0, MemOS, and Hy-Memory, and that EverMemOS was held out of training. arXiv:2610.22196, version 1, was submitted 15 October 2026 by Tianxiang Fei and four other authors. The abstract’s comparison, on fixed 20% subsamples, is a 55.49% macro average versus 54.62% for Claude Opus 5.

Scores Tencent publishes

The figures in this section are Tencent’s. They were not rerun here. The Hunyuan site says that on four benchmarks covering chat, work, and code, Hy-Memory ranks first among the open-source frameworks it names: Mem0, MemOS, EverOS, OpenViking, RepoMemory, and Simple RAG. The headline numbers on that page:

Benchmark	What the site says Hy-Memory scores
LoCoMo, chat accuracy	72.6%
MemoryArena, working-memory average	55.31%
CodeMemoryBench, question answering / task resolution	64.2% / 35.6%
SWE-ContextBench, task resolution	28.3%

The site says the 55.31% average is the mean of travel planning at 59.04% on 270 questions and bundled shopping at 51.57% on 149 questions, 10.2 points above the next result, EverOS. A footnote says the chat LoCoMo figure uses 1,527 non-adversarial questions. A chart on the same page puts Hy-Memory’s stored memory at 6.06% of input tokens. The site says a Simple RAG baseline stores raw windows and should not be compared directly on storage.

Those framework scores are separate from the model-writer scores. On a fixed 20% subsample, the model card’s LoCoMo cells for Hy-Memory-30B-A3B are 59.54 on Mem0, 73.28 on Hy-Memory, 62.98 on MemOS, and 73.28 on EverMemOS. The site’s cross-system LoCoMo cell, 67.27, is the mean of those four. The site’s 72.6% is the framework claim for chat memory.

MemHub, and where the chats go

MemHub is the app layer. The Hunyuan page says a user authorizes a linked tool, imports its past sessions, and the built-in SDK stores them as local memory. It lists 10 tools: WorkBuddy, Cursor, Codex, Claude Code, OpenCode, OpenClaw, Qwen Code, Pi, Hermes, and DeepSeek Harness. It says new sessions can keep syncing into the same memory, and that a memory written in one linked tool is then visible to the others. The MemHub README in that repository calls the product local-first, gives the source version as 1.0.2, and says the release targets are Linux and macOS. Windows is not supported. The default data directory is ~/.one-memory.

The privacy file says source text stays in that local directory. Reading a source, sending work to a model, and letting a host app read memory are separate permissions. If the user picks a third-party model, MemHub sends that service the authorized content the task needs, and the provider, scope, and purpose have to be visible first. The file says the product does not capture the screen, an accessibility tree, OCR, the keyboard, the mouse, the clipboard, or background audio. Uninstall leaves the memory in place unless the user confirms a separate delete. A debug log level can write full prompts to local files. The default level does not.

The Hunyuan page links a repository at Tencent-Hunyuan/MemHub. That URL returned a not-found page on 17 October. The MemHub files that did open are the memhub/ directory inside Tencent-Hunyuan/Hy-Memory.

The public pieces are the Apache 2.0 package hy-memory==2.0.2, the 30B-A3B weights, and a local MemHub app that imports other tools’ chats after authorization. The 72.6% LoCoMo line and the other rankings are Tencent’s reported results.