Personalized assistants often keep long conversation histories to remember user preferences, but sending that entire history to an LLM with every request is expensive. A new preprint on arXiv argues that this is also unnecessary. The authors propose a relevance-filtered context that selects only the most pertinent parts of the conversation.
On the PersonaMem-v2 benchmark, which contains 180 questions, the filtered approach used just 5% of the conversation text. With the GPT-5.6-terra model, it matched the performance of passing the full history. The paper does not report exact accuracy numbers in the abstract, so the magnitude of the match is not yet clear.
This is a single preprint, and the abstract is the only information available here. There are no other sources to compare, so the findings should be treated as preliminary until peer review or replication. If the result holds, it could cut costs for personalized LLM systems without sacrificing quality.