- Who it helps
- Researchers and builders evaluating persistent-memory systems.
- Availability
- Tencent publishes a 30B-total, 3B-active model card with Apache 2.0 metadata. A surrounding application still supplies storage and agent behavior.
A dedicated memory-writing component
The evaluation covers fixed subsets and multiple memory frameworks, including a held-out framework. These are author-reported comparisons of a memory component.
Evaluate corrections and stale facts
Evaluate a correction, a stale preference and an irrelevant fact separately. Keep retrieval quality distinct from the correctness of the final answer.
What you can do
- Check the model card’s evaluation subsets before choosing a memory framework for a comparison. Open reference ↗
Where the evidence stops
The sample does not establish reliable personal-assistant behavior or correct deletion in every application. Model inference was not run.
Sources read. Public evidence checked October 18, 2026. The product, external code and reported outcomes were not independently executed.
Sources & further reading
Hugging Face · official
Hy-Memory model card ↗
Read the memory-writer design and fixed-subset evaluation. Reported results are the publisher’s measurements.
Tencent · page checked 2026-10-18
Coverage note: The linked evidence is a bounded sample. Independent execution and current cross-platform discussion remain incomplete.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
HyMemory is a memory-writing model intended to maintain useful context for an agent over time.