Tencent Hy Memory 2.0: Open-Source Memory Stack for AI Agents (Model + Framework + App)

On October 16, 2026, Tencent Hunyuan released Hy Memory 2.0 as a fully open-source three-layer stack: a dedicated memory model (Hy-Memory-30B-A3B), the Hy-Memory framework for storage and recall, and MemHub — a local-first desktop and CLI app that syncs memory across 10+ AI tools including Cursor, Codex, and Claude Code. The framework ranks #1 among open-source frameworks on four memory benchmarks, and the specialized model edges out Claude Opus 5 on average memory-system scores.

What Tencent actually shipped

Hy Memory 2.0 is not just a model — it's a three-layer memory stack, all open-source under Apache 2.0:

1. Model — Hy-Memory-30B-A3B: A mixture-of-experts model with ~3 billion active parameters, trained specifically for memory extraction, conflict resolution, and memory writing. Validated on four memory systems (Hy-Memory, Mem0, MemOS, EverMemOS). Weights on Hugging Face.
2. Engine — Hy-Memory Framework: Manages memory storage, updates, and recall across three scenarios: chat memory, working memory, and coding memory. Integrates via Python SDK (PyPI: hy-memory) with any OpenAI-compatible LLM and embedding endpoint.
3. App — MemHub: Local-first desktop (Mac) and CLI app that connects to 10+ AI tools, imports their session history into a unified local memory, continuously syncs new sessions, and applies plugin-based recall at task start.

Each layer works alone; together they form a complete agent memory solution. The GitHub repository is at github.com/Tencent-Hunyuan/Hy-Memory.

What's new in 2.0: working memory and coding memory

Hy Memory 1.0 (May 2026, not open-sourced) handled chat memory only. Version 2.0 adds two new memory types:

• Working Memory: Remembers where a task stands — goals, settled approaches, interim results, reusable experience, and the conditions under which that experience applies. Before reusing experience, it checks whether those conditions still hold.
• Coding Memory: Remembers why code changed. For long-horizon tasks, each resume identifies what's being worked on, then retrieves relevant decisions, still-applicable experience, and prior work history — fetching exactly the memory that helps the task move forward.
• Chat Memory (upgraded): Groups scattered facts into structured events with time, place, and people; merges duplicates; resolves contradictions before passing context to the model.

Benchmark results (Tencent-reported, 20% subsample)

Model layer: Hy-Memory-30B-A3B

Four memory systems evaluated (Hy-Memory, Mem0, MemOS, EverMemOS):

• Overall average: 55.49 — #1 of 7 models, ahead of Opus 5 (54.62), DeepSeek V4.1 Flash (53.64), Qwen3-30B-A3B (44.37)
• Training gain vs base: +7.8 points (HY3-A3B Base 47.66 → 55.49)
• Generalization: #1 on every system including EverMemOS (unseen in training)
• By metric (LoCoMo, LongMemEval, PersonaMem v2, Travel, Shopping, CMB): Hy-Memory leads or ties on 5 of 6; Opus 5 leads on LoCoMo (70.23% vs 67.27%)

Framework layer: Hy-Memory vs open-source baselines

• MemoryArena (working memory): 55.31% average — 10.2 pp ahead of runner-up EverOS. Group Travel Planner SPS 59.04%, Bundled Shopping PS 51.57%.
• CodeMemoryBench (coding memory): 64.2% QA accuracy, 35.6% task resolution — #1 open-source.
• SWE-ContextBench (coding memory): 28.3% task resolution — #1 open-source.
• LoCoMo (chat memory): 72.6% accuracy, tokens/question cut from 1,159 to 703 (39.3% reduction) — #1 open-source.
• Storage efficiency: Lowest memory storage footprint across all three scenarios among extraction-based frameworks.

MemHub: the user-facing layer

MemHub is a local-first app (CLI + macOS) that solves a practical problem: preferences you explained in Cursor are unknown to Codex; conventions you set in OpenCode must be re-explained in WorkBuddy. MemHub turns your memory into a long-term asset no single tool can lock in.

Workflow:

1. Import history: Authorize connected tools in MemHub to import past sessions. The built-in Hy-Memory SDK distills them into local memory.
2. Continuous sync: New sessions in every tool sync in real time into the same memory; a memory formed in any tool is immediately visible to all others.
3. Plugin-based recall: Connect each AI tool as a plugin. When you start a task in a connected tool, MemHub manages and applies relevant memories as added context.

Supported tools (10): WorkBuddy, Cursor, Codex, Claude Code, OpenCode, OpenClaw, Qwen Code, Pi, Hermes, DeepSeek Harness.

In production at Tencent

Tencent reports Hy-Memory deployed across Tencent Stocks, iWiki, Tencent News, and other products:

• Tencent Stocks AI Assistant: Multi-hop recall up 40 pp (50% → 90%); overall score 64.3 → 79.6 (+23.8% relative); pass rate 64% → 81%; conflict resolution 67% → 83%.
• iWiki CTX Memory Plugin: Memory alone beats full raw history on answer accuracy: 24.7 → 36.7 (+48.6% relative gain).
• Tencent News Zhisheng/Xinwenmei: Average 91.68/100 on long-term preference understanding; 87.8% of cases score above 90.

Integrated and evaluated in: Finance, News, Knowledge Management, Office Agents, Social & Entertainment, Gaming.

How to get started

Three entry points:

1. App (quickest): Install MemHub (CLI or Mac), authorize your tools, import history, done.
2. Framework (for agent developers): pip install hy-memory, integrate the SDK to manage cross-session information, task progress, and experience.
3. Model (for memory-system researchers): Download Hy-Memory-30B-A3B from Hugging Face; works with Hy-Memory framework and plugs into Mem0, MemOS, etc.

Bottom line: Tencent open-sourced a full agent memory stack on October 16, 2026 — model, framework, and app, all Apache 2.0. The specialized 30B-A3B model beats Opus 5 on memory-system averages; the framework leads open-source baselines on four benchmarks; MemHub syncs memory across 10+ tools locally. If you're building agents that need to remember across sessions and tools, this is the stack to evaluate first.

Prices and benchmark figures are standard published rates as of October 24, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.