Description

The Focus agent (arXiv:2601.07190) introduces model-controlled active context compression: the agent itself decides when and what to compact into a persistent Knowledge block, inspired by slime mold exploration strategies. Unlike Zeph's current compaction (threshold-triggered summarization by a separate LLM call), Focus gives the model agency over its own context lifecycle.

Key result: 22.7% token reduction with identical task performance on agentic benchmarks.

Relevance to Zeph

Zeph's current compaction (zeph-core) triggers when context exceeds a token threshold, calling a compaction provider to summarize. The Focus model proposes a complementary approach:

• Model emits a structured <consolidate> action during normal generation
• Agent loop extracts key learnings → persists in a separate Knowledge block
• Raw interaction history is then pruned (not just summarized)
• Result: agent retains what matters, discards the rest — more targeted than blanket summarization

Key Mechanism

1. Agent generates a special consolidation action when it judges history is getting noisy
2. Key learnings from that session window extracted and appended to persistent Knowledge block
3. History window prior to consolidation pruned (not carried forward)
4. Knowledge block is always kept in context; history is rolling

Difference from Current Compaction

Implementation Sketch

• New ConsolidateAction variant in the agent action enum
• ContextBuilder maintains a persistent KnowledgeBlock in addition to history
• Compaction provider used only for extraction (fast model), not full-context summarization
• Config: [context.focus] with enabled, consolidation_provider, knowledge_max_tokens

Estimated Complexity

Medium. Fits within existing compaction infrastructure (~1 sprint to prototype).

Source: https://arxiv.org/abs/2601.07190