CompactionConfig

Configuration for context compaction.

Quick Example

from mamba_agents import CompactionConfig

config = CompactionConfig(
    strategy="hybrid",
    trigger_threshold_tokens=100000,
    target_tokens=80000,
    preserve_recent_turns=10,
    preserve_system_prompt=True,
)

Configuration Options

Option	Type	Default	Description
strategy	str	"sliding_window"	Compaction strategy
trigger_threshold_tokens	int	100000	Token count to trigger
target_tokens	int	80000	Target after compaction
preserve_recent_turns	int	10	Recent turns to keep
preserve_system_prompt	bool	True	Always keep system
summarization_model	str	"same"	Model for summaries

Available Strategies

• sliding_window - Remove oldest messages
• summarize_older - LLM summarization
• selective_pruning - Remove tool pairs
• importance_scoring - LLM scoring
• hybrid - Combine strategies

API Reference

CompactionConfig

Bases: BaseModel

Configuration for context window compaction.

ATTRIBUTE	DESCRIPTION
strategy	Compaction strategy to use. TYPE: CompactionStrategyType
trigger_threshold_tokens	Token count that triggers compaction. TYPE: int
target_tokens	Target token count after compaction. TYPE: int
preserve_recent_turns	Number of recent turns to always preserve. TYPE: int
preserve_system_prompt	Always preserve the system prompt. TYPE: bool
summarization_model	Model to use for summarization (or "same"). TYPE: str