context pruning

**Context pruning** is the **selective removal of low-value prompt content to maximize useful information density within limited context windows** - it helps maintain performance as conversations and retrieved data grow. **What Is Context pruning?** - **Definition**: Filtering strategy that drops redundant, stale, or irrelevant tokens before inference. - **Pruning Targets**: Greetings, repeated confirmations, obsolete instructions, and low-salience details. - **Decision Criteria**: Relevance to current task, recency, conflict status, and dependency importance. - **Complementary Methods**: Often combined with summarization and retrieval-based rehydration. **Why Context pruning Matters** - **Token Efficiency**: Frees capacity for high-impact instructions and evidence. - **Latency Improvement**: Smaller prompts reduce response time and compute cost. - **Reasoning Quality**: Less noise improves model focus on active objectives. - **Stability**: Reduces conflicts from outdated or superseded conversation fragments. - **Scalable Memory**: Enables longer sessions without uncontrolled context growth. **How It Is Used in Practice** - **Rule-Based Filters**: Apply deterministic policies for removing routine low-value turns. - **Semantic Scoring**: Rank history snippets by relevance to current user intent. - **Safety Preservation**: Never prune mandatory policy and system-control instructions. Context pruning is **a practical optimization for long-context assistant pipelines** - careful removal of low-value tokens improves cost, speed, and answer relevance without sacrificing critical memory.