Active Context Compression (2026)

Focus is an agent-centric architecture proposed in the paper 'Active Context Compression: Autonomous Memory Management in LLM Agents'. It addresses 'Context Bloat' in LLM agents handling long-horizon software engineering tasks. Inspired by the exploration strategies of slime mold (Physarum polycephalum), the Focus Agent autonomously decides when to consolidate key learnings into a persistent 'Knowledge' block and actively prunes the raw interaction history. This self-regulation reduces token usage while maintaining task accuracy. Evaluated on SWE-bench Lite with Claude Haiku 4.5, Focus achieved a 22.7% token reduction (14.9M to 11.5M tokens) with identical accuracy (60% pass rate). The architecture uses an optimized scaffold with persistent bash and string-replacement editor, performing an average of 6.0 autonomous compressions per task, with token savings up to 57% on individual instances. The approach opens pathways for cost-aware agentic systems without sacrificing performance.