Tencent open-sources Hy Memory 2.0 as a private memory steward

Tencent Hunyuan released Hy-Memory 2.0 on October 16, 2026, as an open-source stack of three layers: a memory model, a framework, and an app called MemHub. The launch page calls that combination a private memory steward, meant to remember what someone said, how a situation developed, where the work stands, and why the code was changed. The layers can be used alone or combined. The official @TencentHunyuan post the same day used the steward line and named the same three pieces — model, framework, and MemHub — for chat, work, and code memories.

Treat the rankings that follow as Tencent’s. Nothing in the announcement is an outside audit.

Three layers, three scenes

Hy-Memory-30B-A3B is the model, trained to read and write memory. The page describes an A3B activation scale, about 3 billion activated parameters, and says the weights can run with Hy-Memory or with other systems, including Mem0 and MemOS.

Hy-Memory is the engine. It manages storage, updates, and recall through an SDK, across chat memory, working memory, and code memory.

MemHub is the app, as a CLI and a Mac app. Tencent says you authorize associated tools, import their past sessions, and have the built-in Hy-Memory SDK deposit that history as local memory. New sessions can sync into the same store, and a memory written in one tool is supposed to be visible in the others. A plugin then recalls related memory as extra context when a task starts.

The same page dates version 1.0 to May 2026. That release covered chat memory only, was not open-sourced, and shipped as a paper and a product. Tencent says the underlying design is in a paper under the name DCPM. Version 2.0 keeps that base — session, tags, and graph — upgrades chat memory, and adds working memory and code memory. The company calls Hy-Memory the first memory framework to propose those two scenes. That first-of-its-kind line is its claim.

In the page’s description, chat memory moves from scattered facts to an event thread: time, place, people, and later developments, with duplicates merged and conflicts stripped before the model sees it. Working memory is meant to hold a task’s goal, settled methods, stage results, and reusable experience, and to apply that experience only when the current case meets the conditions. Code memory is for long tasks: identify which job is underway, then retrieve related decisions, still-valid experience, and the earlier work.

Scores Tencent is publishing

Tencent says that on four benchmarks spanning chat, work, and code, Hy-Memory ranked first among the open-source frameworks it compared: Mem0, MemOS, EverOS, OpenViking, RepoMemory, and Simple RAG. The published figures are LoCoMo at 72.6 percent accuracy, MemoryArena at 55.31 percent average, CodeMemoryBench at 64.2 percent for question answering and 35.6 percent solve rate, and SWE-ContextBench at 28.3 percent solve rate.

On MemoryArena, the page reports 59.04 percent SPS on 270 group-travel-planner items and 51.57 percent PS on 149 bundled-shopping items, averaged to 55.31 percent, and says that leads second-place EverOS by 10.2 percentage points. A working-memory chart labels Hy-Memory’s storage occupancy at 6.06 percent, defined as bank tokens divided by input tokens.

The model test is separate. Tencent says Hy-Memory-30B-A3B scored 55.49, first of seven models, on an equal-weight average of LoCoMo, LongMemEval, PersonaMem v2, Travel, Shopping, and CMB, after runs through Hy-Memory, Mem0, MemOS, and EverMemOS. The page lists Opus5 at 54.62, DeepSeek V4.1 Flash at 53.64, and Qwen3-30B-A3B at 44.37. It also says the model led on each of those four systems, including EverMemOS, which was not seen in training. Hy-Memory, Mem0, and MemOS were used in RL training. A base labeled HY3-A3B Base sits at 47.66, which the page calls a 7.8-point gain after memory-specific training. The official post compresses this to Hy-Memory-30B-A3B beating Opus 5 across major frameworks.

Ten tools, stated as a product claim

The launch page says 10 kinds of AI tools can share one personal memory, and it prints WorkBuddy, Cursor, Codex, Claude Code, OpenCode, OpenClaw, Qwen Code, Pi, Hermes, and DeepSeek Harness. The post says MemHub, on CLI and Mac, syncs across 10+ AI tools. That is Tencent’s product claim, not a verified integration list.

Tencent also says Hy-Memory is already in several of its own businesses, with internal numbers attached. For the 自选股 AI assistant, an A-share research memory eval is reported moving from 64.3 to 79.6 overall and multi-hop recall from 50 percent to 90 percent, with DeepSeek V4 Flash as the memory model and the prior scheme as the baseline. An iWiki plugin test, 251 questions with the same model and prompts, is reported at 36.7 using only Hy-Memory versus 24.7 using the full original dialogue. A Tencent News 智声 offline eval is listed at an average 91.68 out of 100 across 41 complex cases from 10 users.

Start from the path you actually need. Use MemHub if you want the app, the Hy-Memory SDK if you are adding memory to an agent, and the Hy-Memory-30B-A3B weights if you want the read-and-write model. The October 16 launch page links all three, and the same-day post is the short official wording.