Tencent open-sources Hy-Memory dual-system cognitive memory engine and MemHub
Tencent Hunyuan released Hy Memory 2.0 on PyPI and open-sourced the Hy-Memory monorepo with MemHub, a local-first desktop and CLI product for LLM agents. The repository includes the Hy-Memory-30B-A3B model on Hugging Face and is licensed under Apache-2.0.
China context
- Original name
- 腾讯混元
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders outside China can use the hy-memory Python SDK and MemHub to add governed, cross-application memory to their agents without relying on proprietary APIs.
- For investors
- Investors should track adoption of Hy-Memory and MemHub as a signal of Tencent's strategy to capture developer mindshare in the agent tooling market through open-source releases.
Tencent Hunyuan has open-sourced Hy-Memory, a production-grade dual-system cognitive memory engine and Python SDK for self-evolving LLM agents. It provides specialized memory pipelines for Chat, Working, and Coding tasks. MemHub, built on Hy-Memory, is a local-first desktop and CLI product that unifies conversation history and coding traces from user-authorized AI applications such as Claude Code, Cursor, Codex, OpenCode, and Qwen Code into a single managed runtime and local data directory. The repository includes the Hy-Memory-30B-A3B model, a 30B-A3B MoE model trained via attributed reinforcement learning for long-term memory writing, available on Hugging Face. The project is licensed under Apache-2.0.
Hy-Memory implements a dual-process cognitive memory architecture with an 8-layer hierarchy for chat memory, two-layer fact and reusable-experience memory for working, and version-evolution diff-chain tracking for coding. The Hy-Memory-30B-A3B model is a 30B-A3B mixture-of-experts model trained with attributed reinforcement learning for high-precision long-term memory extraction and conflict reconciliation. MemHub provides a local-first runtime that aggregates memory from multiple AI coding assistants, enabling cross-application memory recall.
Developers outside China can now integrate a production-grade memory layer into their LLM agents using an Apache-2.0 licensed SDK and local-first tooling, reducing the cost and complexity of building persistent memory for multi-application AI workflows. This directly competes with proprietary memory solutions from Western AI coding assistants by offering an open, self-hostable alternative.
For enterprises and developers, Hy-Memory and MemHub provide a way to unify and govern memory across multiple AI coding tools, potentially reducing vendor lock-in and improving agent performance on long-horizon tasks. The open-source license allows commercial use without licensing fees.
The next observable signal is whether MemHub gains adoption among developers using Claude Code, Cursor, or Codex, as indicated by GitHub stars, issues, and community contributions. Another signal is the release of the CMB benchmark results or acceptance at ICLR 2027, which would validate the coding memory approach.