Skip to content

v0.52.1: conversation-aware summarization

Choose a tag to compare

@hughpyle hughpyle released this 20 Feb 04:23
· 877 commits to main since this release

Auto-detects conversation content (User:/Assistant: turn patterns) and uses a fact-preserving summarization prompt instead of the document-oriented one.

  • Conversations get 300-word budget (vs 200 for documents) with emphasis on preserving dates, names, locations, and chronological order
  • Zero config: auto-detection via speaker-turn patterns (≥4 markers)
  • No protocol change: existing SummarizationProvider implementations work unchanged
  • Tested on LongMemEval temporal-reasoning benchmark: 4/5 accuracy with 10x context compression