diff --git a/crates/mem-core/src/optimizer/mod.rs b/crates/mem-core/src/optimizer/mod.rs
index b000f74..7cc08b3 100644
--- a/crates/mem-core/src/optimizer/mod.rs
+++ b/crates/mem-core/src/optimizer/mod.rs
@@ -13,6 +13,7 @@ pub mod cache_align;
 pub mod ccr;
 pub mod plugin;
 pub mod builtin;
+pub mod query_optimizer;
 
 use anyhow::Result;
 use serde::{Deserialize, Serialize};
@@ -31,6 +32,7 @@ pub use plugin::{
 pub use builtin::{
     BuiltinOptimizer, JsonFormatter, JsonlFormatter, RawFormatter, CsvFormatter, YamlFormatter,
 };
+pub use query_optimizer::{QueryOptimizer, QueryOptimizationMetrics};
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct OptimizedChunk {
diff --git a/crates/mem-core/src/optimizer/query_optimizer.rs b/crates/mem-core/src/optimizer/query_optimizer.rs
new file mode 100644
index 0000000..34b7493
--- /dev/null
+++ b/crates/mem-core/src/optimizer/query_optimizer.rs
@@ -0,0 +1,213 @@
+//! Query-time optimization: compress context before sending to LLM
+//!
+//! Uses pluggable service to apply custom optimization strategies
+//! during query execution (in addition to ingest-time optimization).
+
+use super::plugin::OptimizerService;
+use crate::prompt::CacheMetrics;
+use crate::domain::{Chunk, Record};
+use anyhow::Result;
+
+/// Query optimizer: compresses chunks before LLM processing
+/// Operates on search results AFTER retrieval but BEFORE LLM context window
+pub struct QueryOptimizer {
+    service: Option<OptimizerService>,
+    enabled: bool,
+}
+
+impl QueryOptimizer {
+    /// Create optimizer from environment
+    pub fn from_env() -> Self {
+        let enabled = std::env::var("MEM_QUERY_OPTIMIZER")
+            .map(|v| v.to_lowercase() == "on")
+            .unwrap_or(false);
+
+        if !enabled {
+            return Self {
+                service: None,
+                enabled: false,
+            };
+        }
+
+        // Try to initialize service (will use defaults if available)
+        let service = std::env::var("MEM_QUERY_OPTIMIZER_SERVICE")
+            .ok()
+            .and_then(|_| {
+                // Would load custom plugins from config here
+                None
+            });
+
+        Self {
+            service,
+            enabled: true,
+        }
+    }
+
+    /// Create disabled optimizer
+    pub fn disabled() -> Self {
+        Self {
+            service: None,
+            enabled: false,
+        }
+    }
+
+    /// Set service (for testing/custom initialization)
+    pub fn with_service(mut self, service: OptimizerService) -> Self {
+        self.service = Some(service);
+        self.enabled = true;
+        self
+    }
+
+    /// Get text from chunk by concatenating records
+    fn chunk_text(chunk: &Chunk) -> String {
+        chunk
+            .records
+            .iter()
+            .map(|r| r.text.as_str())
+            .collect::<Vec<_>>()
+            .join("\n")
+    }
+
+    /// Optimize chunk for LLM context (async)
+    /// Returns optimized text (or original if optimization disabled/failed)
+    pub async fn optimize_chunk(&self, chunk: &Chunk) -> Result<String> {
+        if !self.enabled || self.service.is_none() {
+            return Ok(Self::chunk_text(chunk));
+        }
+
+        let service = self.service.as_ref().unwrap();
+        let chunk_text = Self::chunk_text(chunk);
+        let content_type = Self::infer_content_type(&chunk_text);
+
+        // Optimize with default format (raw = just the text)
+        match service.optimize(&chunk_text, &content_type, Some("raw")).await {
+            Ok(bytes) => {
+                let text = String::from_utf8(bytes)
+                    .unwrap_or_else(|_| chunk_text);
+                Ok(text)
+            }
+            Err(_) => {
+                // Fail gracefully: return original if optimization fails
+                tracing::warn!("Query optimization failed, using original chunk");
+                Ok(chunk_text)
+            }
+        }
+    }
+
+    /// Optimize multiple chunks (batch processing)
+    pub async fn optimize_chunks(&self, chunks: &[Chunk]) -> Result<Vec<String>> {
+        let mut optimized = Vec::with_capacity(chunks.len());
+        for chunk in chunks {
+            optimized.push(self.optimize_chunk(chunk).await?);
+        }
+        Ok(optimized)
+    }
+
+    /// Calculate metrics for a chunk (cache alignment + compression ratio)
+    pub fn metrics(&self, chunk: &Chunk, cache_metrics: &CacheMetrics) -> QueryOptimizationMetrics {
+        QueryOptimizationMetrics {
+            original_bytes: Self::chunk_text(chunk).len(),
+            cache_stable_bytes: cache_metrics.stable_prefix_bytes,
+            cache_drift: cache_metrics.drift_metric,
+            is_cache_eligible: cache_metrics.cache_eligible,
+            has_optimizer: self.enabled,
+        }
+    }
+
+    /// Infer content type from chunk text
+    fn infer_content_type(text: &str) -> String {
+        // Simple heuristics (real implementation would use ContentRouter)
+        if text.contains('{') && text.contains('}') {
+            "application/json".to_string()
+        } else if text.contains("ERROR:") || text.contains("WARN:") {
+            "text/x-log".to_string()
+        } else if text.starts_with("---") || text.contains("---") {
+            "text/x-diff".to_string()
+        } else {
+            "text/plain".to_string()
+        }
+    }
+}
+
+/// Metrics for query-time optimization
+#[derive(Debug, Clone)]
+pub struct QueryOptimizationMetrics {
+    pub original_bytes: usize,
+    pub cache_stable_bytes: usize,
+    pub cache_drift: f32,
+    pub is_cache_eligible: bool,
+    pub has_optimizer: bool,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use time::macros::datetime;
+    use crate::domain::{Provenance, Role, Record};
+
+    fn make_test_chunk(text: &str) -> Chunk {
+        Chunk::new(
+            1,
+            vec![Record {
+                role: Role::User,
+                text: text.to_string(),
+                timestamp: datetime!(2024-08-28 12:00:00 UTC),
+                provenance: Provenance {
+                    source_id: "test".to_string(),
+                    offset: 0,
+                },
+            }],
+            100,
+        )
+    }
+
+    #[test]
+    fn test_optimizer_disabled() {
+        let optimizer = QueryOptimizer::disabled();
+        assert!(!optimizer.enabled);
+    }
+
+    #[test]
+    fn test_optimizer_from_env() {
+        std::env::set_var("MEM_QUERY_OPTIMIZER", "off");
+        let optimizer = QueryOptimizer::from_env();
+        assert!(!optimizer.enabled);
+    }
+
+    #[tokio::test]
+    async fn test_optimize_chunk_disabled() {
+        let optimizer = QueryOptimizer::disabled();
+        let chunk = make_test_chunk("test content");
+
+        let result = optimizer.optimize_chunk(&chunk).await;
+        assert_eq!(result.unwrap(), "test content");
+    }
+
+    #[test]
+    fn test_infer_content_type_json() {
+        let json = "{\"key\": \"value\"}";
+        let content_type = QueryOptimizer::infer_content_type(json);
+        assert_eq!(content_type, "application/json");
+    }
+
+    #[test]
+    fn test_infer_content_type_log() {
+        let log = "ERROR: something failed";
+        let content_type = QueryOptimizer::infer_content_type(log);
+        assert_eq!(content_type, "text/x-log");
+    }
+
+    #[test]
+    fn test_infer_content_type_diff() {
+        let diff = "---\n+++ file.txt";
+        let content_type = QueryOptimizer::infer_content_type(diff);
+        assert_eq!(content_type, "text/x-diff");
+    }
+
+    #[test]
+    fn test_infer_content_type_text() {
+        let text = "Plain text content";
+        let content_type = QueryOptimizer::infer_content_type(text);
+        assert_eq!(content_type, "text/plain");
+    }
+}