Context compression & long-run agent context evaluation

Structured summarization by FactoryAI preserves task-critical context better than aggressive compression approaches used by OpenAI and Anthropic.
Key insights
Context compression must preserve task continuity
A comparative evaluation argued that maximizing compression rate alone is insufficient for long-running agent tasks; preserved details such as file paths, error logs and prior decisions are critical. The study recommended optimizing 'per-task token consumption' over raw compression rate to avoid repeated retrievals, hallucinations or task interruption.
Factory anchored iterative summarization leads in real tasks
FactoryAI’s 'anchored iterative summarization' maintains a structured persistent summary (intent, file edits, decisions, next steps) merged incrementally; in large-scale tests it outperformed OpenAI and Anthropic on accuracy and context awareness. The evaluation used probe categories (Recall, Artifact, Continuation, Decision) and GPT-5.2 as an automated judge across production message datasets.
Trade-offs: OpenAI compression vs Anthropic structure
The comparison found OpenAI offered the highest compression ratio (~99.3%) but lost technical specificity (e.g., file paths), while Anthropic produced detailed structured summaries but re-generated them each time, risking drift. All methods struggled with artifact tracking, suggesting the need for supplemental indexing.