---
id: "2026-10-15-tencent-hy-memory-30b-a3b-memory-writer"
url: "https://postcutoff.com/e/2026-10-15-tencent-hy-memory-30b-a3b-memory-writer/"
as_of: "2026-10-17T07:29:00+02:00"
date: "2026-10-15"
date_precision: day
category: open-source
importance: 2
confidence: high
status: [Confirmed]
sources: 4
editor: human editor
human_review: null
version: null
---

As of: 2026-10-17 07:29 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/e/2026-10-15-tencent-hy-memory-30b-a3b-memory-writer/

# Tencent open-sources Hy-Memory-30B-A3B memory writer

Full title: Tencent Hunyuan open-sources Hy-Memory 2.0 and Hy-Memory-30B-A3B, a small 'memory writer' model for agent memory frameworks

Tencent Hunyuan released Hy-Memory-30B-A3B (Apache 2.0, ~3B active parameters) on Hugging Face on Oct 15, 2026, with the Hy-Memory framework (chat, working and code memory) and a tech report (arXiv 2610.22196). The model replaces the general-purpose LLM that writes long-term memory inside frameworks such as Mem0, MemOS and Hy-Memory. In Tencent's cross-framework test it averages 55.49, slightly above Claude Opus 5 (54.62) and DeepSeek V4.1 Flash (53.64) as the memory writer.

## Key facts
- Model: Hy-Memory-30B-A3B, MoE with ~3B active parameters, English and Chinese, Apache 2.0 (Hugging Face card)
- Role: executes the write path of external memory frameworks; drop-in replacement for the LLM inside Mem0, MemOS and Hy-Memory (chat, working and code memory); EverMemOS held out of training
- Training: two stages inside the deployed harnesses: on-policy distillation from a teacher on student-written state chains, then RL with chunk-level credit over memory banks
- Tencent's results (fixed 20% subsample; six metrics incl. LoCoMo, LongMemEval, PersonaMem-v2 and a code-memory benchmark): average 55.49 vs Claude Opus 5 54.62, DeepSeek V4.1 Flash 53.64, Hy3 51.63, Qwen3.6-35B-A3B 50.37; only the memory writer is swapped, answer model and judge held fixed
- Pandaily (Oct 17) reported it as Hy Memory 2.0 with new working and code memory and a MemHub, Apache 2.0

## What happened

Agent memory systems usually call a large general model to decide what to store. Tencent trained a small dedicated model for that job and
released it with its own framework. The comparison numbers are Tencent's and were not independently checked.

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 168 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 107 days after its cutoff
- Gemini 4 Argon (training cutoff March 2026): 198 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 137 days after its cutoff

## Sources

1. [Hugging Face: tencent/Hy-Memory-30B-A3B](https://huggingface.co/tencent/Hy-Memory-30B-A3B) (huggingface.co, code)
2. [GitHub: Tencent-Hunyuan/Hy-Memory](https://github.com/Tencent-Hunyuan/Hy-Memory) (github.com, code)
3. [Tech report: arXiv 2610.22196](https://arxiv.org/pdf/2610.22196) (arxiv.org, paper)
4. [Pandaily: Tencent Hunyuan open-sources Hy Memory 2.0, adding working and code memory for AI agents](https://pandaily.com/tencent-hunyuan-hy-memory-2-0-open-source-working-code-memory-memhub-apache-2) (pandaily.com, press)

## Changes

- 2026-10-17 (filed): Created (Hugging Face card read)