Tencent Hunyuan open-sources Hy Memory 2.0 agent memory framework and model

Summary

Tencent Hunyuan has released Hy Memory 2.0 as open source. This three-layered agent memory stack consists of a specialized model with approximately 3 billion parameters, a framework managing chat/work/code memory, and a local MemHub application. The company claims it achieved the highest scores in various memory benchmarks and narrowly outperformed Claude Opus 5 on a fixed 20% subsample.

Why it matters

Persistent and structured memory is one of the biggest practical gaps in today's agent ecosystems. A fully open-source stack comprising an SDK, model weights, and a desktop application lowers the barrier for developers looking to build agents that maintain context during long, multi-tool tasks. If the benchmark claims are verified through independent replication, this offers the open-source community a reliable alternative to proprietary memory layers like Mem0 or MemOS.

Highlights

• Hy-Memory-30A3B (~3B active parameters) scored 55.49 points on a fixed 20% benchmark subsample across four evaluation frameworks, while Claude Opus 5 scored 54.62 points.
• The framework claims to be a first mover in the 'working memory' and 'code memory' categories, outperforming Mem0, MemOS, EverOS, and others in four benchmarks.
• The MemHub application (CLI + Mac) synchronously aggregates chat history from over 10 AI tools into a single local memory repository via the Hy Memory SDK.

Critical take

The comparison with Opus 5 relies on a 0.87-point difference in a 20% subsample chosen by the company. This test pits a 3-billion-parameter specialized model against a much larger general-purpose model, which hardly counts as a fair one-on-one contest. Furthermore, the benchmarks are entirely self-reported by Tencent, with no reference to third-party replications. Given prior research on episodic and procedural memory in agent systems, the 'first in industry' framing for work/code memory is also debatable.