AI agents forget. Close the session and your preferences, project setup and half your instructions are gone. Tencent's answer, released on 15 October 2026, is not a bigger model — it's a specialist: Hy Memory 2.0, an open-source stack whose only job is deciding what an agent should remember.
What's in the release
Three layers:
- Hy-Memory-30B-A3B — a 30B-parameter MoE model (~3B active) purpose-trained for the "memory-write" step: what to keep, how to store it, and how new facts overwrite old ones. Apache 2.0, on Hugging Face.
- The framework — a Python SDK (pip install hy-memory) that splits memory into three types instead of one chat log: chat memory, working memory, and coding memory — diff chains that keep why a change was made, not just what changed.
- MemHub — a local-first app that syncs authorised history across your AI tools, so one agent can recall what another did. No Windows build yet, and the day-one installer had gaps.
About those benchmarks
Tencent reports a 55.49% macro average across memory benchmarks vs 54.62% for Claude Opus 5. Read that as parity, not a rout — these are Tencent's own benchmarks, not independently replicated. The solid facts are the license, the tiny inference footprint, and that nothing else open-source is purpose-built for this slot: the model drops into Mem0 or MemOS as the memory-writer without changing your storage or retrieval.
The bigger shift
October's releases keep unbundling the monolith: DeepSeek's V4.1-Flash for speed, Meta's Muse Wonder splitting memory and focus, and now Tencent offloading remembering entirely. A 3B-active specialist doing the bookkeeping means your frontier model spends its tokens thinking.
Full breakdown with pros, cons and the honest caveats: Hy Memory 2.0 on NeurAITool
Top comments (0)