Unverified60% confidenceBenchmarkExact time
In memory tests, retaining the full context allowed the model to accurately locate specific details mentioned by students 95% of the time, whereas after summary compression the accuracy dropped sharply to 32%.
2
Sources
60%
Confidence
Medium-term (~90 days)
Relevance
8/20/2026
First Seen
Valid until: 11/18/2026
Sources
Related Claims
UnverifiedAccording to the ForgetEval research, when leading memory systems were scored on test cases written by people other than the authors themselves, internal test scores of 63%-68% plummeted to 28%-33%.71% similarUnverifiedWhen recombining rules seen during training, symbolic descriptions match or even exceed the example approach, reaching 99% versus 95% accuracy on new combinations.71% similarUnverifiedExperimental data shows that when key information is located at the beginning and end of the input sequence, the model's retrieval accuracy can be as high as over 90%, but when located in the middle, it may drop sharply to below 60%.70% similarUnverifiedAccording to the STALE research, most dedicated memory frameworks score below 10% on the task of recognizing the failure of their own memory.69% similarUnverifiedThe FEB-30 benchmark proposed in the arXiv paper "Beyond Downstream QA" uses an inversion paradigm to construct deterministic ground truth, and tests show that a representative memory system correctly classifies memory operations (create, update, strengthen, supersede) only 20% of the time.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/778558API
curl https://kongchang.com/api/v1/knowledge/claims/778558MCP
get_claim(id=778558)