Unverified50% confidenceFactExact time
After compressing the conversation with /compact, context usage in the demo dropped from 24% to 13%; after clearing with /clear, 10% was still occupied by the system prompt, tools, and CLAUDE.md.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/17/2026
First Seen
Valid until: 10/31/2026
Sources
Related Entities
Related Claims
Unverified研究表明在超过8-10轮对话后，模型对初始系统提示中规则的遵循率可下降20%-40%71% similarUnverifiedCalls containing multiple tools rose to 47% within three days after the rule was written, then fell back to 18%, and then dropped to 10%.69% similarUnverified语义缓存机制可将实际API调用量降低15%-40%68% similarUnverified实验数据显示，经过5-10轮迭代训练后，模型在低频词汇和长尾知识上的覆盖率下降超过40%，而输出文本在表面流畅性上几乎没有可察觉的退化68% similarUnverifiedModule residual anchoring lowered the hit rate under natural questioning from 28.7% to 22.0% while also reducing the measured behavioral drift, a trade-off of exchanging recall for preservation of non-target behavior.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1050955API
curl https://kongchang.com/api/v1/knowledge/claims/1050955MCP
get_claim(id=1050955)