Lost in the Middle & Position Effects

Core

Manage conversation context to preserve critical information across long interactions · Difficulty 3/5

lost-in-middlecontextattentionposition-effects

Prerequisites

Large language models attend more reliably to information at the beginning and end of their context, with reduced attention to middle sections -- the "lost in the middle" effect.

The Problem

When aggregated results total ~75K tokens:

First ~15K tokens: reliably cited (primacy effect)

Last ~10K tokens: reliably cited (recency effect)

Middle ~50K tokens: frequently omitted, even when containing critical findings

Solutions

Key findings summary at the top: Place a condensed summary of all critical information at the beginning (leverages primacy effect)

Explicit section headers: Help the model navigate and attend to middle content

Structured data: Use key facts, citations, and relevance scores instead of verbose content

Reduce total volume: Have upstream agents return structured data rather than verbose content and reasoning chains

Anti-Pattern: Streaming Sequentially

Processing results one source at a time prevents holistic cross-source analysis and doesn't solve the attention distribution issue.

Key Takeaways

✓Models attend best to beginning and end of context, less to the middle
✓Place key findings summary at the beginning of large contexts
✓Use section headers and structured data to improve middle-context attention

Related Concepts

Progressive Summarization Risks

Extract critical transactional facts into a persistent block outside summarized history

Context Budget Management & Upstream Reduction

Reduce data volume at the source rather than trying to handle large inputs downstream

Test Yourself1 of 2

When the synthesis agent processes aggregated results from all subagents (~75K tokens total), it reliably cites findings from the first and last sections but frequently omits critical findings from the middle sections. What's the most effective fix?