Tencent's Hy-Memory-30B-A3B beats Claude Opus 5 at agent memory writing with 55.49% avg
dair_ai · x · 2026-10-18
dairai highlights Tencent Hunyuan's Hy-Memory-30B-A3B, a 3B-active model trained solely for long-term memory writing—deciding what an agent stores, how precisely, and how new evidence updates existing state.
- One model works across Mem0, MemOS, and Hy-Memory, replacing framework-specific LLM writers
- Ranks first on all four evaluated harnesses (including held-out EverMemOS) on fixed 20% benchmark subsamples, with a 55.49% macro average vs 54.62% for Claude Opus 5
- Two-stage training: on-policy distillation for each framework's write protocol, then RL rewarding writes that later prove answerable, with over-storage penalties
- Paper on arXiv (2610.22196); weights and code open-sourced on Hugging Face and GitHub
More from coding & agent
- Dev open-sources an X network map tool with a Grok agent prompt to collect data — jacob_posel · 2026-10-18
- Serverless, explained with a pizza shop and ghost kitchen analogy — thisiskp_ · 2026-10-18
- Pydantic AI CodeMode speculatively runs tools mid-generation, cutting 3 lookups from 8.7s to 4.7s — samuelcolvin · 2026-10-18
- 18 AI engineering concepts, from RAG variants to LoRA/QLoRA, mapped to open-source repos — MaryamMiradi · 2026-10-18
- Bilibili Creators Used Claude to Make an Undertale Music Video, Surprising Reddit — Luciaka · 2026-10-18
- Model-agnostic harnesses will never get the most out of a model — gethackteam · 2026-10-18