Context Management
Can You Delete What the User Said?
AI output can be summarized, tool results can be truncated, but once the user's original words are deleted, they're gone forever. And the user will always notice: "I clearly said xxx — how did you forget?"
Comparison: Consequences of Two Compression Approaches
Option A: Delete the User's Messages
Option B: Compress Only the AI's Output
User Messages Are Sacred
User Messages = Sacred
Every sentence the user inputs carries needs, preferences, and context. Once deleted, the AI doesn't just lose information — it loses the user's trust: "If you can't even remember what I said, what's the point of this AI?"
Compression Priority
1
Delete First: Raw Tool/Function Call Output
Search results, API JSON responses — once processed, they serve no further purpose
2
Compress Next: Long AI Responses
What the AI says can be replaced with a summary — just keep the core conclusions
3
Never Touch: The User's Original Messages
What the user said, what they asked for, their preferences — all of this is sacred
User messages are sacred: better to delete 1,000 words of AI output than to touch 10 words from the user. Compression priority from highest to lowest: tool output > AI responses > user messages (never touch).