One of the most significant technical innovations in Claude Opus 4.6 is context compaction—a technology designed to solve the persistent problem of "context rot" that has plagued long-running AI conversations. The results are striking: 76% accuracy on the MRCR v2 benchmark compared to just 18.5% for the previous generation.

Understanding Context Rot

Context rot refers to the degradation of response quality and relevance as conversations grow longer. Even with large context windows, AI models have historically struggled to maintain coherent references to information mentioned early in extended exchanges.

This manifests in several ways:

  • Forgetting details mentioned earlier in the conversation
  • Contradicting previous statements
  • Losing track of the conversation's overall purpose
  • Failing to connect related information across long exchanges

The MRCR v2 Benchmark

The Multi-turn Response Consistency and Relevance (MRCR) v2 benchmark specifically tests an AI's ability to maintain coherent, contextually-aware responses over extended conversations. It evaluates both consistency with prior statements and relevance to accumulated context.

How Context Compaction Works

Context compaction employs a sophisticated approach to managing conversation history. Rather than treating all context equally, the system:

1. Intelligent Summarization

Earlier portions of conversations are automatically summarized, preserving key facts, decisions, and contextual information while reducing token count. This isn't simple truncation—it's semantic compression that maintains information fidelity.

2. Importance Weighting

The system identifies which parts of the conversation are most likely to be referenced later, giving them higher priority in the compacted representation. User corrections, explicit instructions, and key data points receive special attention.

3. Dynamic Expansion

When the model detects that it needs to reference compacted content, it can selectively expand relevant portions back to full detail, ensuring accuracy when specific information is needed.

Benchmark Results

The improvement in MRCR v2 scores tells a compelling story:

  • Claude Sonnet 4.5: 18.5% accuracy
  • Claude Opus 4.6: 76% accuracy
  • Improvement: 4.1x better performance

This dramatic improvement makes Claude Opus 4.6 particularly well-suited for:

  • Long-running research assistants
  • Multi-session project collaborations
  • Customer support bots with persistent context
  • Coding assistants working on large projects

Technical Implementation

For developers, context compaction is handled automatically—no special configuration required. However, the API does expose some controls:

  • Compaction threshold: Configure when compaction begins
  • Priority markers: Flag specific content as high-priority
  • Expansion hints: Suggest topics that may need full context

Looking Ahead

Anthropic has indicated that context compaction represents the first step in a broader initiative to make AI systems more capable of long-term, coherent interaction. Future developments may include cross-session memory and user-specific learning patterns.

Source: InfoQ →