Hy Memory 2.0: Tencent's Open-Source Model That Only Remembers
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)…
On 15 October 2026, Tencent introduced Hy Memory 2.0, an open-source model designed to focus solely on managing an AI agent's memory. This release aims to solve the issue of AI agents forgetting information when a session is closed by offering a three-layer solution: Hy-Memory-30B-A3B, Apache 2.0 licensed and hosted on Hugging Face, along with a Python SDK for easy integration.
The framework divides memory into three distinct types: chat memory, working memory, and coding memory, with diff chains tracking the reasons behind changes rather than just the changes themselves. MemHub, a local-first application, ensures authorized history is shared across various AI tools, allowing different agents to access each other's memories.
Tencent's benchmarks show a 55.49% macro average in memory benchmarks, slightly outperforming Claude Opus 5's 54.62%. The release's key strengths include a permissive license, a small inference footprint, and the absence of competing open-source models specifically designed for memory management. This unification of memory management within a specialized model allows larger frontier models to allocate more tokens towards thinking, rather than remembering.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
