Tencent Hunyuan has released version 2.0 of Hy Memory, its memory framework for AI agents, and open-sourced all three layers of the stack: a dedicated memory model, the framework itself and a local memory app called MemHub. The layers can be used separately or combined. Tencent says Hy Memory is already used by internal teams in finance, news, knowledge management, office agents, social entertainment and games.
The framework handles storage, updating and evolution of three memory types: chat, working and code. Developers connect it to agents through an SDK, published on PyPI as hy-memory 2.0.2. Working memory and code memory are new in 2.0. They record a task's goal, decisions already made, intermediate results and reusable lessons along with the conditions under which those lessons apply. Before resuming a long-running task, the system first works out which task is in progress, then retrieves relevant decisions, still-valid experience and prior work history. Chat memory now groups scattered facts into events with time, place and people, merging duplicates and resolving conflicts before passing context to the model.
Tencent reports first place against open-source frameworks including Mem0, MemOS and EverOS on four benchmarks. By its figures, working memory scored an average 55.31% on MemoryArena travel-planning and shopping tasks. Code memory reached 64.2% question-answering accuracy and 35.6% task resolution on CodeMemoryBench, and 28.3% task resolution on SWE-ContextBench. Chat memory hit 72.6% accuracy on LoCoMo while cutting tokens per question from 1,159 to 703, a 39.3% reduction.
The model layer is Hy-Memory-30B-A3B, a mixture-of-experts model with about 3 billion active parameters trained for memory extraction and conflict handling. Its technical report evaluates it on a fixed 20% subsample of the benchmarks across four memory systems. On that subsample it posts a macro average of 55.49, slightly above Claude Opus 5 at 54.62 and DeepSeek V4.1 Flash at 53.64. Weights are on Hugging Face.
MemHub is a local-first app with a command-line tool and a Mac client. Once a user authorizes connected tools, it imports their past sessions into a single local memory store and keeps syncing new ones. Tencent says it currently plugs into 10 AI tools, including WorkBuddy, Cursor, Codex, Qwen Code, OpenCode and DeepSeek Harness.
The GitHub repository's license file states that Hy-Memory is released under Apache 2.0.
SEE ALSO: Tencent Hunyuan Releases Hy4 Preview · Doubao Work Adds One-Click Import of Skills, Memory and Chat History