2016 String Algorithm Adapted to Fix LLM Memory Loss in Multi-Document Summarization

Large language models struggle to accurately summarize many documents under strict length limits, either ignoring middle content or forgetting earlier information — biases known as primacy and recency bias. A developer revisited a hierarchical string-concatenation method he published in 2016, originally designed to reduce memory overhead in Google Apps Script, and adapted it for AI document aggregation. The approach, called Pyramid Aggregation, processes documents in a tree-like structure rather than all at once or one by one, distributing attention more evenly across all inputs. When tested against traditional batch and sequential methods, the technique achieved a 95% speedup in processing time while significantly reducing information loss. The findings have been published as a preprint on Zenodo under the title 'Pyramid Aggregator: Mitigating Information Loss in Multi-Document Fact Extraction via Hierarchical Merging'.