research(memory): density-aware context compression — adaptive ratio selection by information density (arXiv:2603.25926)

[bug-ops]

Source

arXiv:2603.25926 — Density-aware Soft Context Compression with Semi-Dynamic Compression Ratio

Finding

Adaptive context compression that varies compression ratio based on turn-level information density rather than uniform compression. Uses a jointly-trained Discrete Ratio Selector that classifies each turn as high/medium/low density and applies different compression budgets accordingly.

Zeph applicability

• Current CompactionStrategy::Reactive applies uniform summarization threshold (compaction_threshold = 0.80)
• Tool outputs (code, JSON, file content) are high-density and compress poorly at the same ratio as conversational turns
• Implementation sketch: pre-classify each message by density heuristic (token count, code block ratio, JSON ratio); apply tighter budget to low-density turns; preserve more of high-density tool outputs; integrate with ACON guidelines (compression_guidelines feature)