Tencent Hunyuan released the entire Hy Memory 2.0 stack for AI agent long-term memory under the Apache-2.0 license on October 15, 2026: the Python SDK hy-memory==2.0.1 on PyPI, the Hy-Memory monorepo on GitHub, and the Hy-Memory-30B-A3B model on Hugging Face. The model is built as a 30B-parameter MoE with 3B active parameters and was trained specifically for long-term memory writing; according to the authors, it ranks first on all four memory frameworks and outperforms Claude Opus 5. In addition to the model, the release includes a framework with three memory scenes and the MemHub application, which synchronizes context between different AI assistants, so you only need to describe it once.

image
image
image

What Happened

On October 15, 2026, the Tencent Hunyuan team released three artifacts simultaneously. The Python SDK hy-memory==2.0.1 appeared on PyPI; the Hy-Memory monorepo opened on GitHub, containing the MemHub application as a CLI and Mac app; the Hy-Memory-30B-A3B model weights were uploaded to Hugging Face under Apache-2.0. The model itself is a Mixture of Experts with 30B parameters, of which 3B are active; it was trained via on-policy distillation and attributed reinforcement learning with chunk-level GRPO, specifically for the long-term memory writing operation. The authors claim first place on all four memory frameworks — Mem0, Hy-Memory, MemOS, and EverMemOS — with an average result of 55.49% compared to 54.62% for Claude Opus 5. The framework manages three memory scenes: chat with an L0–L7 level hierarchy, working with current session facts and reusable experience, and coding with a diff version chain. MemHub synchronizes memory across 11 AI tools, including Claude Code, Cursor, Codex, OpenCode, and Qwen Code.

Context

Long-term memory is one of the bottlenecks in agentic systems: a general-purpose LLM itself remembers nothing between sessions, so external fact stores based on frameworks like Mem0, MemOS, and EverMemOS are attached to it, and the writing is usually performed by the same large model following instructions from a prompt. Hy Memory 2.0 offers a different architecture: memory writing is separated into a distinct trainable component with its own specialized model, and the framework defines what and at what level is remembered — a level hierarchy in chat, working facts and reusable experience in sessions, and version chains in code. This is methodologically sound because long-term memory writing is a specific operation: you need to decide what is worth saving, at what level of the hierarchy to place it, and how to reuse it later, and a universal prompt handles such decisions poorly. The choice of comparison point is also telling: the frontier Claude Opus 5, meaning the bet is on the effect of specialization rather than simply increasing model size. Memory is thus following the same path by which retrievers and tool calls have already separated into independent layers in agentic systems.

Why This Matters for the Industry

The main signal for the industry is not a specific benchmark number, but the architectural decision: long-term memory is being formalized as an independent infrastructure layer with its own model, framework, and synchronization protocol between tools. An open MoE model with 3B active parameters that can realistically be run locally, the Apache-2.0 license, and claimed compatibility with popular frameworks Mem0 and MemOS reduce the cost and barrier to embedding long-term memory into existing agentic stacks. For memory startups, this is commoditization: memory writing becomes cheap at the level of open infrastructure, a pure memory subscription as a business comes under pressure, and value shifts to governance, observability, and integrations on top of the stack. If independent measurements confirm the advantage of a compact specialized model over general-purpose LLMs, this will become an argument for narrow specialized small MoE models for individual agentic functions — memory, planning, verification — and for a standard component of agentic platforms in the form of a shared memory service for multiple agents; the emergence of a separate product category around memory hubs is also likely.

Why This Matters for Users

The stack is available now: just run pip install hy-memory==2.0.1, download the Hy-Memory-30B-A3B weights from Hugging Face, and launch MemHub, which is distributed as a CLI and Mac application. The practical meaning for the reader: project context is described once, after which it is used by Claude Code, Cursor, Codex, OpenCode, Qwen Code, and other supported tools — without repeated explanations to each assistant. Due to the 3B active parameters, local deployment looks more realistic than one might expect from the 30B figure, so experiments are possible on your own hardware without the cloud. A reasonable first step is a prototype and pilot in a day: spin up a memory service next to your current agent and A/B compare it with your existing memory layer on your own traffic before judging the quality of writing.

What Is Still Unknown / Limitations

The key numbers — an average of 55.49% compared to 54.62% for Claude Opus 5 — are self-reported: the comparison was published by the release authors, and one of the four frameworks, Hy-Memory, is their own, so the identity of baseline configurations and memory pipelines requires separate verification. There are no ablations in the available materials that would separate the contribution of distillation, reinforcement learning, and ordinary prompting of the same model, so the specialization effect is currently a hypothesis. Engineering data important for production — latency, throughput, hardware requirements, behavior under quantization — is not provided in the sources. Finally, the availability of artifacts is provable by downloading, but the quality of memory writing and the viability of the "describe context once — works everywhere" pattern have not yet been confirmed by independent measurements, so the stack should be evaluated on your own data.

Sources

Author

Look at AI, editorial team