// Generated from the feature catalog
Context Window Compression
Picks the Right Brain for the Job · Shipped · workstation, home
Level-of-Detail memory management keeping long conversations from blowing the context ceiling.
As conversations grow, raw token counts can exceed the model’s context window. The compression subsystem applies lingua and semantic compressors to distill earlier turns into denser representations while preserving critical facts and tool outputs.
The lingua compressor removes syntactic filler and redundancy. The semantic compressor identifies and merges duplicate information across turns. Together they maintain a sliding window of full-fidelity recent turns backed by compressed summaries of older context.
Limits and invariants
Section titled “Limits and invariants”Compression never discards tool outputs or system-state observations — only conversational filler and redundant explanations are candidates for reduction.
Where this lives
Section titled “Where this lives”halbert_core/halbert_core/compression/lingua_compressor.pyhalbert_core/halbert_core/compression/semantic_compressor.py