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

Explanation

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 Yourself

1 / 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?