腾讯 Hy Memory 2.0：面向 AI Agent 的开源记忆写入模型

Tencent's Hy Memory 2.0: An Open-Source Memory Writer for AI Agents Every AI agent has the same embarrassing problem: it forgets. Bigger context windows never fixed it — they just made forgetting more expensive. Tencent Hunyuan's answer, released October 15, 2026, is refreshingly direct: stop asking the chatbot to remember things, and train a model whose entire job is writing memories. The centerpiece is Hy-Memory-30B-A3B: roughly 30 billion total parameters with only about 3 billion active per token (a mixture-of-experts design). It is not a chatbot — it is the librarian. It reads a conversation or coding trace and decides what to retain, how precisely to represent it, and how new evidence updates the persistent store. Why the write path matters Most agent-memory systems today (Mem0, MemOS, and similar) work the same way: after each turn, a general-purpose language model reads the exchange and decides what to file away. That general model is expensive, slow, and was trained to chat, not to file. This is also where memory fails hardest — research shows leading systems return the correct current state only 13–20% of the time after facts change over time. Hy-Memory-30B-A3B is trained against exactly that failure, in two stages: Memory On-Policy Distillation (a frozen Hy3 teacher supervises the student on the student's own output, so supervision matches deployment reality) and Attributed Memory Reinforcement Learning (rewards for how well future questions get answered from the written banks, with credit assigned per chunk). Three layers, one Apache 2.0 drop Hy Memory 2.0 landed as three releases within about 24 hours: The model — weights on Hugging Face under Apache 2.0, a drop-in replacement for the general-purpose LLM inside Mem0, MemOS, and Tencent's own framework. The framework — a Python SDK (hy-memory on PyPI) handling three "memory scenes": chat, working-task facts, and code changes (tracking why a change was made, not just what). Tencent calls it the first framework to explicitly separate working memory from code memory. MemHub — a local-first desktop app (macOS/Linux) that puts authorized conversation traces into one shared local store, so your coding assistants stop starting from zero. The numbers — read carefully Tencent reports a macro average of 55.49 vs Claude Opus 5's 54.62 vs DeepSeek V4.1 Flash's 53.64 across four memory harnesses, and 72.6% on the LoCoMo conversational-memory benchmark. The honest caveat: these are vendor-run, unreplicated benchmarks with a sub-one-point margin. Test-worthy, not trust-worthy. Should you try it? For developers building agents, the drop-in path into Mem0 or MemOS makes a one-week experiment cheap — and the Apache 2.0 license means no legal friction. Two practical notes: official weights exist only in BF16 (~60 GB), though community GGUF quantizations (~18 GB Q4) fit a single 24 GB card; and MemHub's launch was rough (the npm installer it pointed to didn't exist), so budget for a source checkout. The real story here isn't the benchmark table. It's cost: a purpose-trained memory writer at 3B-model serving cost, replacing the frontier chat model doing your memory filing today. Memory is becoming infrastructure — and Tencent just made it free. This is a condensed developer explainer. The full breakdown — model architecture, training stages, complete benchmark tables, pros and cons, and three ways to run it — is on NeurAITool: Tencent's Hy Memory 2.0: Open-Source AI That Never Forgets.