mediar-ai Context Degradation Patterns
Recognize and mitigate context failures as context length increases.
Prerequisites
- Understanding of attention mechanisms
- Familiarity with context windows
Instructions
Degradation Patterns
Lost-in-Middle: Information in center receives less attention.
- 10-40% lower recall for middle content
- Place critical info at beginning or end
Context Poisoning: Errors compound through repeated reference.
- Tool outputs, retrieved docs, or summaries introduce errors
- Creates feedback loops reinforcing incorrect beliefs
Context Distraction: Irrelevant info overwhelms relevant content.
- Single irrelevant document reduces performance
- Models must attend to everything provided
Context Confusion: Model cannot determine which context applies.
- Responses address wrong aspect of query
- Tool calls appropriate for different task
Context Clash: Accumulated information directly conflicts.
- Multi-source retrieval with contradictions
- Version conflicts between outdated and current info
Model Degradation Thresholds
| Model | Degradation Onset | Severe |
|---|---|---|
| GPT-5.2 | ~64K tokens | ~200K |
| Claude Opus 4.5 | ~100K tokens | ~180K |
| Claude Sonnet 4.5 | ~80K tokens | ~150K |
| Gemini 3 Pro | ~500K tokens | ~800K |
Four-Bucket Mitigation
- Write: Save context outside window (scratchpads, files)
- Select: Pull relevant context through retrieval/filtering
- Compress: Reduce tokens via summarization
- Isolate: Split across sub-agents with fresh contexts
Guidelines
- Monitor context length and performance correlation
- Place critical information at beginning or end
- Implement compaction before degradation becomes severe
- Validate retrieved documents for accuracy
- Segment tasks to prevent confusion across objectives
Notes
- RULER benchmark: Only 50% of 32K+ models maintain performance at 32K
- Shuffled haystacks can outperform coherent ones
- Larger context windows do not uniformly improve performance
Source: muratcankoylan/Agent-Skills-for-Context-Engineering