research(memory): Focus autonomous context compression — Knowledge block consolidation, 23% token reduction on SWE-bench (arXiv:2601.07190)

[bug-ops]

Source

arXiv:2601.07190 — "Active Context Compression: Autonomous Memory Management in LLM Agents"

Key Findings

• "Focus" architecture: agents autonomously consolidate interaction history into a persistent Knowledge block, pruning less relevant context
• ~23% token reduction on SWE-bench with no task performance loss
• Biologically inspired exploration/exploitation balance for deciding what to compress vs retain
• Unlike full compaction, only less-relevant segments are compressed — selective and cheaper

Gap in Zeph

Zeph currently uses compress_context in always_on mode — full compaction applied to the entire context window. The Focus approach is more selective: it identifies low-relevance segments and consolidates only those into the Knowledge block, leaving high-relevance recent context intact.

Existing compression issues (#2481 density-aware, #2460 performance-floor) focus on ratio selection, not on selective segment targeting. Focus is architecturally distinct.

Implementation Sketch for Zeph

1. In zeph-memory, add a KnowledgeBlock segment type alongside the existing compaction pipeline
2. Add a relevance scorer (cheap model) to identify segments below a threshold
3. Route low-relevance segments to consolidation into the Knowledge block instead of full compaction
4. Expose [memory.compression] strategy = "focus" | "always_on" config option
5. Surface compression activity in TUI with a spinner ("Consolidating context…")

Agents with equivalent feature

• SWE-bench top performers use selective context management (Cursor shadow workspace, Aider repo-map pruning)

References

• arXiv:2601.07190: https://arxiv.org/abs/2601.07190
• Related open issues: research(memory): density-aware context compression — adaptive ratio selection by information density (arXiv:2603.25926) #2481 (density-aware compression), research(memory): performance-floor context compression — predictor-driven ratio selection (arXiv:2603.19733) #2460 (performance-floor ratio), feat(memory): wire goal_text from agent context into remember() calls for CraniMem write gate #2483 (goal-conditioned write gate)