Pandaily reported that Tencent Hunyuan has released Hy Memory 2.0 and open-sourced its memory model, framework and local app MemHub. The three components can be used separately or together.
The framework now handles working and code memory alongside chat memory. It records a task’s goal, previous decisions, intermediate results and lessons that may be useful again. When a task resumes, the system identifies it and retrieves relevant material. Chat memory instead organises scattered details into events, combining duplicates and resolving conflicts. Developers can connect agents through the hy-memory 2.0.2 SDK on PyPI.
Planning a trip could draw on earlier decisions when a task resumes, if those decisions had been recorded, rather than relying on the conversation alone.
On LoCoMo, Tencent reports 72.6% accuracy for chat memory and a reduction in tokens per question from 1,159 to 703, or 39.3%. Its Hy-Memory-30B-A3B model uses a mixture-of-experts design with about 3 billion active parameters. On a fixed 20% subsample of benchmarks across four memory systems, it scored a macro average of 55.49.
MemHub has a command-line tool and a Mac client that import authorised tools’ past sessions into a local store and continue syncing new ones. The GitHub repository’s licence file lists Apache 2.0 for Hy-Memory, Pandaily reported.
