`tencent-hunyuan/hy-memory` is a dual-system cognitive memory engine for its own self-evolving LLM agents. The latest release shown in the repository is v1.0.5, but the provided information contains no changelog for this version, so it is not possible to confirm what was newly added or fixed. The README summary introduces the project as a Python SDK and explains that it places separate memory pipelines for three work scenarios: chat (Chat), working (Working), and coding (Coding).
Chat memory handles raw conversation transcripts, atomic facts, summaries, schemas, prospective intentions, and more across an 8-level hierarchy, performing conflict reconciliation and cross-domain schema induction. Working memory stores facts and reusable experiences separately, while coding memory preserves code modification trajectories as diff chains between versions. Retrieval uses a path that combines BM25 and embeddings. The README also introduces MemHub, a separate product. It is a desktop and CLI product that unifies the conversation logs and coding traces of AI applications approved by the user into a management runtime and a local data directory.
This architecture does not stop at simply storing memories as vectors; it is an approach that aims to handle the formation of memory and conflict resolution as well. However, the provided materials contain neither the release changes for v1.0.5 nor any performance verification results. The license is also listed as NOASSERTION, so the usage conditions cannot be judged from this information alone. Therefore, it is more accurate to view this as a project worth examining for its memory structure rather than for specific performance or usage rights.