Megadose The AI curve, compressed.

LMLineage: Content Lineage and Context Governance for LLM Agents

· ArXiv · AI/CL/LG ·
LMLineage identifies context an agent has stopped using, then compacts it while preserving user instructions verbatim.

The system links each output sentence to supporting input passages with confidence scores, without access to model internals or additional inference passes. The authors evaluated attribution on 71 annotated sentences from 10 Chinese web-QA cases and 90 turns across three generator models. They report attribution in one second per turn on a single GPU, with stable performance through 40 turns. Drift detection produced no false compactions in their evaluation, though its ability to find unused context was limited by attribution accuracy.

ArXiv · AI/CL/LG's note

score 5

Categories: Research