Core concepts
Segments split long history into topical units.
MemoGrafter detects topic drift and stores contiguous message ranges as topic segments.
Why segments exist
A segment is a range of messages that belong to one topic. It keeps long conversations from becoming one undifferentiated memory blob.
Think of segments as the boundary between raw chat and graph memory. They preserve where a topic started and ended, which makes later recall, Studio inspection, and graft provenance easier to explain.
example
messages 0-4 -> Japan travel planning
messages 5-8 -> cover letter writingDrift signals
MemoGrafter uses embedding distance and optional intent signals to decide when a conversation has moved to a new topic. Short acknowledgements and filler turns are dampened so they do not accidentally create noisy segments.
Embedding distance from the current topic.
Structural phrases like
by the way or going back to.Short-message dampening for filler turns.
Optional LLM ambiguity detection and reentry detection.