Context Compression: Economic Token Pruning
Reducing token costs by summarizing history while keeping logic intact.
Steps
- Summarize older turns once context exceeds 50% of the window.
- Protect 'Hard Facts' (entities/dates) from being lost in summaries.
- Use 'Recursive Summarization' for 100k+ token sessions.
- Compare 'Summary Accuracy' against raw history periodically.
- Drop 'Low-Utility' turns (e.g., greetings, filler) from the context.