Tencent open-sources Hunyuan Hy Memory 2.0, a 30B MoE memory writer for AI agents

Tencent's Hunyuan team has open-sourced Hy Memory 2.0, a toolkit for giving AI agents durable memory across sessions. According to a dev.to writeup, the release landed on October 15, 2026, and its centerpiece is Hy-Memory-30B-A3B: a roughly 30-billion-parameter mixture-of-experts model in which only about 3 billion parameters are active per token, trained to do one job — writing memories rather than holding a conversation.

A specialist for the write path

The dev.to article frames the weakness in today's agent-memory stacks. Systems such as Mem0 and MemOS typically let a general-purpose language model read each exchange after the fact and decide what to file away. That model is costly, slow and trained for conversation rather than record-keeping, and it is where memory fails hardest: research cited in the article says leading systems return the correct current state of facts only 13 to 20 percent of the time once those facts have shifted over a conversation.

Hy-Memory-30B-A3B is meant to replace that general model in the pipeline. It reads a conversation or coding trace and decides what to keep, at what level of detail, and how new evidence should revise the persistent store. Tencent positions it as a drop-in swap inside Mem0, MemOS and its own framework.

Two training stages aimed at forgetting

Per the writeup, training attacked that failure mode directly. A first stage, Memory On-Policy Distillation, has a frozen Hy3 teacher supervise the student on the student's own outputs, so supervision matches what the model will actually produce once deployed. A second stage, Attributed Memory Reinforcement Learning, rewards the model according to how well future questions can be answered from the memory banks it wrote, with credit assigned per chunk.

Three releases under Apache 2.0

The launch arrived as three components within about a day:

• The model: weights on Hugging Face under an Apache 2.0 license.
• The framework: a Python SDK called hy-memory on PyPI, covering three "memory scenes" — chat, working-task facts, and code changes, the last recording why a change was made rather than only what changed. Tencent says it is the first framework to draw an explicit line between working memory and code memory.
• MemHub: a local-first desktop app for macOS and Linux that gathers authorized conversation traces into one shared local store, so coding assistants do not begin every session with no history.

Vendor-run benchmarks, thin margins

Tencent reports a macro average of 55.49 across four memory harnesses, ahead of Claude Opus 5 at 54.62 and DeepSeek V4.1 Flash at 53.64, along with 72.6 percent on the LoCoMo conversational-memory benchmark. The dev.to author's own caveat deserves repeating: these are vendor-run, unreplicated numbers, and the lead is under a single point.

Practical rough edges

Two notes for anyone trying it now. Official weights ship only in BF16, roughly 60 GB, though community GGUF quantizations around 18 GB at Q4 reportedly fit a single 24 GB graphics card. MemHub's debut was also bumpy — the npm installer the launch pointed to reportedly did not exist, so expect to build from source.

Why it matters

The economics are the real story. Memory writing has until now been a side job for a frontier chat model, billed at frontier prices. A purpose-trained writer with roughly 3 billion active parameters can do the same filing for a fraction of the serving cost, and the Apache 2.0 license removes legal friction. The benchmark claims still need independent replication, but the drop-in path into Mem0 or MemOS makes a one-week trial cheap. If the results hold, Hy Memory 2.0 marks a broader shift: memory stops being a feature bolted onto agents and becomes cheap, swappable infrastructure.