Keep It InMind: Benchmarking the Implicit-Association Blind Spot in Agent Memory
Agent memory systems can recall stored facts directly, yet fail when a query only needs them through an implicit link.
The paper introduces InMind, a 125-task benchmark for cases where the relevant memory does not look like the incoming question. In the authors’ tests, the backbone model answered 84.0% of indirect queries when the decisive memory was already in context. When retrieval had to surface that same memory, six memory systems reached at most 14.4%, despite recalling the facts on demand at up to 100%. The authors locate the failure in the query-conditioned retrieval interface and frame routing as the open problem. HF Daily Papers' note
The paper introduces InMind, a 125-task benchmark for cases where the relevant memory does not look like the incoming question. In the authors’ tests, the backbone model answered 84.0% of indirect queries when the decisive memory was already in context. When retrieval had to surface that same memory, six memory systems reached at most 14.4%, despite recalling the facts on demand at up to 100%. The authors locate the failure in the query-conditioned retrieval interface and frame routing as the open problem. HF Daily Papers' note
score 5