An agent hears an instrument again under changed acoustics. It must recognise the recurring source, update what changed and retain what did not. One recall score mixes all three jobs.
AMemOps, posted October 13, evaluates latest-attribute retrieval, unchanged-attribute preservation and absent-source rejection in long audio streams. Its paired controls examine prior exposure, conflicting history and the contribution of erasure. Read the research paper.
Updating is not simply adding another record
A memory that keeps every observation can retrieve an obsolete attribute. A memory that overwrites the entire source can destroy still-valid information. Selective revision requires knowing which property changed.
Prior training exposure also affects interpretation. Recognising a familiar instrument under new conditions is different from maintaining a newly encountered source over time.
For audio assistants, include source identity changes, contradictory attributes and long retrieval delays. Ask both the latest value and an unchanged value after the update.
The benchmark's controlled streams do not cover every noisy real-world recording. They do make the evaluation more diagnostic. A useful audio memory should demonstrate not only that something was heard, but that the right part of what was heard has replaced the corresponding old belief.
Questions, news tips or corrections? Contact [email protected].
