Hy Memory 2.0
Tencent Hunyuan open-sourced Hy Memory 2.0, an agent memory stack with an SDK, a 30B-A3B memory model and the MemHub local app, adding working and code memory.
Version 2.0 adds working memory and code memory to the existing chat memory. Working and code memory record a task's goal, decisions, intermediate results and reusable lessons, and the system identifies the task in progress before retrieving them. Chat memory now groups facts into events and merges duplicates and conflicts. All three layers (model, framework, MemHub) are open-sourced and can be used separately.
- Date
- Saturday 17 October 2026
- Lab
- Tencent Hunyuan
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| LoCoMo accuracy (chat memory) | 72.6% tokens per question cut from 1,159 to 703, a 39.3% reduction | company |
| MemoryArena average (working memory) | 55.31% travel-planning and shopping tasks | company |
| CodeMemoryBench QA accuracy | 64.2% task resolution 35.6% | company |
| SWE-ContextBench task resolution | 28.3% | company |
| Hy-Memory-30B-A3B macro average | 55.49 on a fixed 20% subsample of benchmarks; Claude Opus 5 scored 54.62 and DeepSeek V4.1 Flash 53.64 | company |
Apache 2.0 per the GitHub license file; SDK on PyPI as hy-memory 2.0.2; MemHub plugs into 10 AI tools. Benchmark results are Tencent's own.
Sources
This record was checked against its sources on 18 October 2026. How we check