diff --git a/Cargo.lock b/Cargo.lock
index 794a880..69f1673 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -3655,7 +3655,7 @@ checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a"
 
 [[package]]
 name = "vestige-core"
-version = "1.5.0"
+version = "1.6.0"
 dependencies = [
  "chrono",
  "directories",
@@ -3689,7 +3689,7 @@ dependencies = [
 
 [[package]]
 name = "vestige-mcp"
-version = "1.5.0"
+version = "1.6.0"
 dependencies = [
  "anyhow",
  "axum",
diff --git a/Cargo.toml b/Cargo.toml
index 63ddea3..92c981b 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -7,7 +7,7 @@ members = [
 ]
 
 [workspace.package]
-version = "1.5.0"
+version = "1.6.0"
 edition = "2024"
 license = "AGPL-3.0-only"
 repository = "https://github.com/samvallad33/vestige"
diff --git a/crates/vestige-core/Cargo.toml b/crates/vestige-core/Cargo.toml
index 7312e93..59e486f 100644
--- a/crates/vestige-core/Cargo.toml
+++ b/crates/vestige-core/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "vestige-core"
-version = "1.5.0"
+version = "1.6.0"
 edition = "2024"
 rust-version = "1.85"
 authors = ["Vestige Team"]
diff --git a/crates/vestige-core/src/advanced/adaptive_embedding.rs b/crates/vestige-core/src/advanced/adaptive_embedding.rs
index 91877b7..d4b39d3 100644
--- a/crates/vestige-core/src/advanced/adaptive_embedding.rs
+++ b/crates/vestige-core/src/advanced/adaptive_embedding.rs
@@ -31,13 +31,11 @@
 use serde::{Deserialize, Serialize};
 use std::collections::HashMap;
 
-/// Default embedding dimensions (BGE-base-en-v1.5: 768d, upgraded from MiniLM 384d)
-/// 2026 GOD TIER UPGRADE: +30% retrieval accuracy
-pub const DEFAULT_DIMENSIONS: usize = 768;
+/// Default embedding dimensions after Matryoshka truncation (768 → 256)
+pub const DEFAULT_DIMENSIONS: usize = 256;
 
-/// Code embedding dimensions (when using code-specific models)
-/// Now matches default since we upgraded to 768d
-pub const CODE_DIMENSIONS: usize = 768;
+/// Code embedding dimensions (matches default after Matryoshka truncation)
+pub const CODE_DIMENSIONS: usize = 256;
 
 /// Supported programming languages for code embeddings
 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq, Hash)]
diff --git a/crates/vestige-core/src/embeddings/local.rs b/crates/vestige-core/src/embeddings/local.rs
index 30a5415..9b67629 100644
--- a/crates/vestige-core/src/embeddings/local.rs
+++ b/crates/vestige-core/src/embeddings/local.rs
@@ -18,9 +18,10 @@ use std::sync::{Mutex, OnceLock};
 // CONSTANTS
 // ============================================================================
 
-/// Embedding dimensions for the default model (nomic-embed-text-v1.5)
-/// 768 dimensions with Matryoshka support (can truncate to 256/512 if needed)
-pub const EMBEDDING_DIMENSIONS: usize = 768;
+/// Embedding dimensions after Matryoshka truncation
+/// Truncated from 768 → 256 for 3x storage savings with only ~2% quality loss
+/// (Matryoshka Representation Learning — the first N dims ARE the N-dim representation)
+pub const EMBEDDING_DIMENSIONS: usize = 256;
 
 /// Maximum text length for embedding (truncated if longer)
 pub const MAX_TEXT_LENGTH: usize = 8192;
@@ -277,7 +278,7 @@ impl EmbeddingService {
             ));
         }
 
-        Ok(Embedding::new(embeddings[0].clone()))
+        Ok(Embedding::new(matryoshka_truncate(embeddings[0].clone())))
     }
 
     /// Generate embeddings for multiple texts (batch processing)
@@ -307,7 +308,7 @@ impl EmbeddingService {
                 .map_err(|e| EmbeddingError::EmbeddingFailed(e.to_string()))?;
 
             for emb in embeddings {
-                all_embeddings.push(Embedding::new(emb));
+                all_embeddings.push(Embedding::new(matryoshka_truncate(emb)));
             }
         }
 
@@ -338,6 +339,26 @@ impl EmbeddingService {
 // SIMILARITY FUNCTIONS
 // ============================================================================
 
+/// Apply Matryoshka truncation: truncate to EMBEDDING_DIMENSIONS and L2-normalize
+///
+/// Nomic Embed v1.5 supports Matryoshka Representation Learning,
+/// meaning the first N dimensions of the 768-dim output ARE a valid
+/// N-dimensional embedding with minimal quality loss (~2% on MTEB for 256-dim).
+#[inline]
+pub fn matryoshka_truncate(mut vector: Vec<f32>) -> Vec<f32> {
+    if vector.len() > EMBEDDING_DIMENSIONS {
+        vector.truncate(EMBEDDING_DIMENSIONS);
+    }
+    // L2-normalize the truncated vector
+    let norm = vector.iter().map(|x| x * x).sum::<f32>().sqrt();
+    if norm > 0.0 {
+        for x in &mut vector {
+            *x /= norm;
+        }
+    }
+    vector
+}
+
 /// Compute cosine similarity between two vectors
 #[inline]
 pub fn cosine_similarity(a: &[f32], b: &[f32]) -> f32 {
diff --git a/crates/vestige-core/src/embeddings/mod.rs b/crates/vestige-core/src/embeddings/mod.rs
index 2dec019..fcbb9d1 100644
--- a/crates/vestige-core/src/embeddings/mod.rs
+++ b/crates/vestige-core/src/embeddings/mod.rs
@@ -14,8 +14,8 @@ mod hybrid;
 mod local;
 
 pub use local::{
-    cosine_similarity, dot_product, euclidean_distance, Embedding, EmbeddingError,
-    EmbeddingService, BATCH_SIZE, EMBEDDING_DIMENSIONS, MAX_TEXT_LENGTH,
+    cosine_similarity, dot_product, euclidean_distance, matryoshka_truncate, Embedding,
+    EmbeddingError, EmbeddingService, BATCH_SIZE, EMBEDDING_DIMENSIONS, MAX_TEXT_LENGTH,
 };
 
 pub use code::CodeEmbedding;
diff --git a/crates/vestige-core/src/search/hybrid.rs b/crates/vestige-core/src/search/hybrid.rs
index bfaf908..33c492a 100644
--- a/crates/vestige-core/src/search/hybrid.rs
+++ b/crates/vestige-core/src/search/hybrid.rs
@@ -117,8 +117,8 @@ pub struct HybridSearchConfig {
 impl Default for HybridSearchConfig {
     fn default() -> Self {
         Self {
-            keyword_weight: 0.5,
-            semantic_weight: 0.5,
+            keyword_weight: 0.3,
+            semantic_weight: 0.7,
             rrf_k: 60.0,
             min_semantic_similarity: 0.3,
             source_limit_multiplier: 2,
diff --git a/crates/vestige-core/src/search/reranker.rs b/crates/vestige-core/src/search/reranker.rs
index 55d58ca..5f1058a 100644
--- a/crates/vestige-core/src/search/reranker.rs
+++ b/crates/vestige-core/src/search/reranker.rs
@@ -1,14 +1,17 @@
 //! Memory Reranking Module
 //!
-//! ## GOD TIER 2026: Two-Stage Retrieval
+//! ## Two-Stage Retrieval with Cross-Encoder
 //!
-//! Uses fastembed's reranking model to improve precision:
-//! 1. Stage 1: Retrieve top-50 candidates (fast, high recall)
-//! 2. Stage 2: Rerank to find best top-10 (slower, high precision)
+//! Uses fastembed's Jina Reranker v1 Turbo (38M params) cross-encoder
+//! for high-precision reranking:
+//! 1. Stage 1: Retrieve top-50 candidates via hybrid search (fast, high recall)
+//! 2. Stage 2: Cross-encoder rerank to find best top-10 (slower, high precision)
 //!
-//! This gives +15-20% retrieval precision on complex queries.
+//! Falls back to BM25-like term overlap scoring when the cross-encoder
+//! model is unavailable.
 
-// Note: Mutex and OnceLock are reserved for future cross-encoder model implementation
+#[cfg(feature = "embeddings")]
+use fastembed::{RerankInitOptions, RerankerModel, TextRerank};
 
 // ============================================================================
 // CONSTANTS
@@ -83,21 +86,15 @@ impl Default for RerankerConfig {
     }
 }
 
-/// Service for reranking search results
+/// Service for reranking search results using a cross-encoder model
 ///
-/// ## Usage
-///
-/// ```rust,ignore
-/// let reranker = Reranker::new(RerankerConfig::default());
-///
-/// // Get initial candidates (fast, recall-focused)
-/// let candidates = storage.hybrid_search(query, 50)?;
-///
-/// // Rerank for precision
-/// let reranked = reranker.rerank(query, candidates, 10)?;
-/// ```
+/// When the `embeddings` feature is enabled and `init_cross_encoder()` is called,
+/// uses Jina Reranker v1 Turbo for neural cross-encoder scoring.
+/// Falls back to BM25-like term overlap when the model is unavailable.
 pub struct Reranker {
     config: RerankerConfig,
+    #[cfg(feature = "embeddings")]
+    cross_encoder: Option<TextRerank>,
 }
 
 impl Default for Reranker {
@@ -108,24 +105,61 @@ impl Default for Reranker {
 
 impl Reranker {
     /// Create a new reranker with the given configuration
+    ///
+    /// The cross-encoder model is NOT loaded here — call `init_cross_encoder()`
+    /// explicitly to load it. This keeps construction fast and test-friendly.
     pub fn new(config: RerankerConfig) -> Self {
-        Self { config }
+        Self {
+            config,
+            #[cfg(feature = "embeddings")]
+            cross_encoder: None,
+        }
+    }
+
+    /// Initialize the cross-encoder model (Jina Reranker v1 Turbo, ~150MB)
+    ///
+    /// Downloads the model on first call. Call this during server startup,
+    /// NOT in tests or hot paths.
+    #[cfg(feature = "embeddings")]
+    pub fn init_cross_encoder(&mut self) {
+        if self.cross_encoder.is_some() {
+            return; // Already initialized
+        }
+
+        let options = RerankInitOptions::new(RerankerModel::JINARerankerV1TurboEn)
+            .with_show_download_progress(true);
+
+        match TextRerank::try_new(options) {
+            Ok(model) => {
+                eprintln!("[vestige] Cross-encoder reranker loaded (Jina Reranker v1 Turbo)");
+                self.cross_encoder = Some(model);
+            }
+            Err(e) => {
+                eprintln!("[vestige] Cross-encoder unavailable, using BM25 fallback: {e}");
+            }
+        }
+    }
+
+    /// Check if the cross-encoder model is available
+    pub fn has_cross_encoder(&self) -> bool {
+        #[cfg(feature = "embeddings")]
+        {
+            self.cross_encoder.is_some()
+        }
+        #[cfg(not(feature = "embeddings"))]
+        {
+            false
+        }
     }
 
     /// Rerank candidates based on relevance to the query
     ///
-    /// This uses a cross-encoder model for more accurate relevance scoring
-    /// than the initial bi-encoder embedding similarity.
-    ///
-    /// ## Algorithm
-    ///
-    /// 1. Score each (query, candidate) pair using cross-encoder
-    /// 2. Sort by score descending
-    /// 3. Return top-k results
+    /// Uses cross-encoder model when available for neural relevance scoring.
+    /// Falls back to BM25-like term overlap scoring otherwise.
     pub fn rerank<T: Clone>(
-        &self,
+        &mut self,
         query: &str,
-        candidates: Vec<(T, String)>, // (item, text content)
+        candidates: Vec<(T, String)>,
         top_k: Option<usize>,
     ) -> Result<Vec<RerankedResult<T>>, RerankerError> {
         if query.is_empty() {
@@ -138,15 +172,43 @@ impl Reranker {
 
         let limit = top_k.unwrap_or(self.config.result_count);
 
-        // For now, use a simplified scoring approach based on text similarity
-        // In a full implementation, this would use fastembed's RerankerModel
-        // when it becomes available in the public API
+        // Try cross-encoder first
+        #[cfg(feature = "embeddings")]
+        if let Some(ref mut model) = self.cross_encoder {
+            let documents: Vec<&str> = candidates.iter().map(|(_, text)| text.as_str()).collect();
+
+            if let Ok(rerank_results) = model.rerank(query, &documents, false, None) {
+                let mut results: Vec<RerankedResult<T>> = rerank_results
+                    .into_iter()
+                    .filter_map(|rr| {
+                        candidates.get(rr.index).map(|(item, _)| RerankedResult {
+                            item: item.clone(),
+                            score: rr.score,
+                            original_rank: rr.index,
+                        })
+                    })
+                    .collect();
+
+                results.sort_by(|a, b| {
+                    b.score.partial_cmp(&a.score).unwrap_or(std::cmp::Ordering::Equal)
+                });
+
+                if let Some(min_score) = self.config.min_score {
+                    results.retain(|r| r.score >= min_score);
+                }
+
+                results.truncate(limit);
+                return Ok(results);
+            }
+            // Cross-encoder failed on this call — fall through to BM25 fallback
+        }
+
+        // Fallback: BM25-like scoring
         let mut results: Vec<RerankedResult<T>> = candidates
             .into_iter()
             .enumerate()
             .map(|(rank, (item, text))| {
-                // Simple BM25-like scoring based on term overlap
-                let score = self.compute_relevance_score(query, &text);
+                let score = Self::compute_relevance_score(query, &text);
                 RerankedResult {
                     item,
                     score,
@@ -155,25 +217,19 @@ impl Reranker {
             })
             .collect();
 
-        // Sort by score descending
         results.sort_by(|a, b| b.score.partial_cmp(&a.score).unwrap_or(std::cmp::Ordering::Equal));
 
-        // Apply minimum score filter
         if let Some(min_score) = self.config.min_score {
             results.retain(|r| r.score >= min_score);
         }
 
-        // Take top-k
         results.truncate(limit);
 
         Ok(results)
     }
 
-    /// Compute relevance score between query and document
-    ///
-    /// This is a simplified BM25-inspired scoring function.
-    /// A full implementation would use a cross-encoder model.
-    fn compute_relevance_score(&self, query: &str, document: &str) -> f32 {
+    /// BM25-inspired term overlap scoring (fallback when cross-encoder unavailable)
+    fn compute_relevance_score(query: &str, document: &str) -> f32 {
         let query_lower = query.to_lowercase();
         let query_terms: Vec<&str> = query_lower.split_whitespace().collect();
         let doc_lower = document.to_lowercase();
@@ -184,22 +240,19 @@ impl Reranker {
         }
 
         let mut score = 0.0;
-        let k1 = 1.2_f32; // BM25 parameter
-        let b = 0.75_f32; // BM25 parameter
-        let avg_doc_len = 500.0_f32; // Assumed average document length
+        let k1 = 1.2_f32;
+        let b = 0.75_f32;
+        let avg_doc_len = 500.0_f32;
 
         for term in &query_terms {
-            // Count term frequency
             let tf = doc_lower.matches(term).count() as f32;
             if tf > 0.0 {
-                // BM25-like term frequency saturation
                 let numerator = tf * (k1 + 1.0);
                 let denominator = tf + k1 * (1.0 - b + b * (doc_len / avg_doc_len));
                 score += numerator / denominator;
             }
         }
 
-        // Normalize by query length
         if !query_terms.is_empty() {
             score /= query_terms.len() as f32;
         }
@@ -223,7 +276,7 @@ mod tests {
 
     #[test]
     fn test_rerank_basic() {
-        let reranker = Reranker::default();
+        let mut reranker = Reranker::default();
 
         let candidates = vec![
             (1, "The quick brown fox".to_string()),
@@ -234,13 +287,12 @@ mod tests {
         let results = reranker.rerank("fox", candidates, Some(2)).unwrap();
 
         assert_eq!(results.len(), 2);
-        // Results with "fox" should be ranked higher
         assert!(results[0].item == 1 || results[0].item == 3);
     }
 
     #[test]
     fn test_rerank_empty_candidates() {
-        let reranker = Reranker::default();
+        let mut reranker = Reranker::default();
         let candidates: Vec<(i32, String)> = vec![];
 
         let results = reranker.rerank("query", candidates, Some(5)).unwrap();
@@ -249,7 +301,7 @@ mod tests {
 
     #[test]
     fn test_rerank_empty_query() {
-        let reranker = Reranker::default();
+        let mut reranker = Reranker::default();
         let candidates = vec![(1, "some text".to_string())];
 
         let result = reranker.rerank("", candidates, Some(5));
@@ -258,22 +310,28 @@ mod tests {
 
     #[test]
     fn test_min_score_filter() {
-        let reranker = Reranker::new(RerankerConfig {
+        let mut reranker = Reranker::new(RerankerConfig {
             min_score: Some(0.5),
             ..Default::default()
         });
 
         let candidates = vec![
-            (1, "fox fox fox".to_string()),  // High relevance
-            (2, "completely unrelated".to_string()),  // Low relevance
+            (1, "fox fox fox".to_string()),
+            (2, "completely unrelated".to_string()),
         ];
 
         let results = reranker.rerank("fox", candidates, None).unwrap();
 
-        // Only high-relevance results should pass the filter
         assert!(results.len() <= 2);
         if !results.is_empty() {
             assert!(results[0].score >= 0.5);
         }
     }
+
+    #[test]
+    fn test_default_has_no_cross_encoder() {
+        let reranker = Reranker::default();
+        // Default constructor does NOT load the model — fast and test-friendly
+        assert!(!reranker.has_cross_encoder());
+    }
 }
diff --git a/crates/vestige-core/src/search/vector.rs b/crates/vestige-core/src/search/vector.rs
index 35b6db0..5ada474 100644
--- a/crates/vestige-core/src/search/vector.rs
+++ b/crates/vestige-core/src/search/vector.rs
@@ -17,9 +17,9 @@ use usearch::{Index, IndexOptions, MetricKind, ScalarKind};
 // CONSTANTS
 // ============================================================================
 
-/// Default embedding dimensions (BGE-base-en-v1.5: 768d)
-/// 2026 GOD TIER UPGRADE: +30% retrieval accuracy over MiniLM (384d)
-pub const DEFAULT_DIMENSIONS: usize = 768;
+/// Default embedding dimensions after Matryoshka truncation (768 → 256)
+/// 3x storage savings with only ~2% quality loss on MTEB benchmarks
+pub const DEFAULT_DIMENSIONS: usize = 256;
 
 /// HNSW connectivity parameter (higher = better recall, more memory)
 pub const DEFAULT_CONNECTIVITY: usize = 16;
@@ -137,7 +137,7 @@ impl VectorIndex {
         let options = IndexOptions {
             dimensions: config.dimensions,
             metric: config.metric,
-            quantization: ScalarKind::F32,
+            quantization: ScalarKind::F16,
             connectivity: config.connectivity,
             expansion_add: config.expansion_add,
             expansion_search: config.expansion_search,
@@ -325,7 +325,7 @@ impl VectorIndex {
         let options = IndexOptions {
             dimensions: config.dimensions,
             metric: config.metric,
-            quantization: ScalarKind::F32,
+            quantization: ScalarKind::F16,
             connectivity: config.connectivity,
             expansion_add: config.expansion_add,
             expansion_search: config.expansion_search,
diff --git a/crates/vestige-core/src/storage/sqlite.rs b/crates/vestige-core/src/storage/sqlite.rs
index 143f4e6..4d9733a 100644
--- a/crates/vestige-core/src/storage/sqlite.rs
+++ b/crates/vestige-core/src/storage/sqlite.rs
@@ -22,10 +22,10 @@ use crate::memory::{
 use crate::search::sanitize_fts5_query;
 
 #[cfg(feature = "embeddings")]
-use crate::embeddings::{Embedding, EmbeddingService, EMBEDDING_DIMENSIONS};
+use crate::embeddings::{matryoshka_truncate, Embedding, EmbeddingService, EMBEDDING_DIMENSIONS};
 
 #[cfg(feature = "vector-search")]
-use crate::search::{reciprocal_rank_fusion, VectorIndex};
+use crate::search::{linear_combination, VectorIndex};
 
 // ============================================================================
 // ERROR TYPES
@@ -202,7 +202,13 @@ impl Storage {
 
         for (node_id, embedding_bytes) in embeddings {
             if let Some(embedding) = Embedding::from_bytes(&embedding_bytes) {
-                if let Err(e) = index.add(&node_id, &embedding.vector) {
+                // Handle Matryoshka migration: old 768-dim → truncate to 256-dim
+                let vector = if embedding.dimensions != EMBEDDING_DIMENSIONS {
+                    matryoshka_truncate(embedding.vector)
+                } else {
+                    embedding.vector
+                };
+                if let Err(e) = index.add(&node_id, &vector) {
                     tracing::warn!("Failed to load embedding for {}: {}", node_id, e);
                 }
             }
@@ -690,7 +696,7 @@ impl Storage {
             }
             #[cfg(all(feature = "embeddings", feature = "vector-search"))]
             SearchMode::Hybrid => {
-                let results = self.hybrid_search(&input.query, input.limit, 0.5, 0.5)?;
+                let results = self.hybrid_search(&input.query, input.limit, 0.3, 0.7)?;
                 results.into_iter().map(|r| r.node).collect()
             }
             #[cfg(not(all(feature = "embeddings", feature = "vector-search")))]
@@ -1257,7 +1263,7 @@ impl Storage {
         };
 
         let combined = if !semantic_results.is_empty() {
-            reciprocal_rank_fusion(&keyword_results, &semantic_results, 60.0)
+            linear_combination(&keyword_results, &semantic_results, keyword_weight, semantic_weight)
         } else {
             keyword_results.clone()
         };
diff --git a/crates/vestige-mcp/Cargo.toml b/crates/vestige-mcp/Cargo.toml
index 0f1ce4d..a66cdef 100644
--- a/crates/vestige-mcp/Cargo.toml
+++ b/crates/vestige-mcp/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "vestige-mcp"
-version = "1.5.0"
+version = "1.6.0"
 edition = "2024"
 description = "Cognitive memory MCP server for Claude - FSRS-6, spreading activation, synaptic tagging, and 130 years of memory research"
 authors = ["samvallad33"]
diff --git a/crates/vestige-mcp/src/cognitive.rs b/crates/vestige-mcp/src/cognitive.rs
index 7feb33e..34e04af 100644
--- a/crates/vestige-mcp/src/cognitive.rs
+++ b/crates/vestige-mcp/src/cognitive.rs
@@ -64,7 +64,7 @@ pub struct CognitiveEngine {
 impl CognitiveEngine {
     /// Initialize all cognitive modules with default configurations.
     pub fn new() -> Self {
-        Self {
+        let engine = Self {
             // Neuroscience
             activation_network: ActivationNetwork::new(),
             synaptic_tagging: SynapticTaggingSystem::new(),
@@ -98,6 +98,8 @@ impl CognitiveEngine {
             // Search
             reranker: Reranker::new(RerankerConfig::default()),
             temporal_searcher: TemporalSearcher::new(),
-        }
+        };
+
+        engine
     }
 }
diff --git a/crates/vestige-mcp/src/dashboard/handlers.rs b/crates/vestige-mcp/src/dashboard/handlers.rs
index 17b3f93..ffb11be 100644
--- a/crates/vestige-mcp/src/dashboard/handlers.rs
+++ b/crates/vestige-mcp/src/dashboard/handlers.rs
@@ -38,7 +38,7 @@ pub async fn list_memories(
         {
             // Use hybrid search
             let results = storage
-                .hybrid_search(query, limit, 0.5, 0.5)
+                .hybrid_search(query, limit, 0.3, 0.7)
                 .map_err(|_| StatusCode::INTERNAL_SERVER_ERROR)?;
 
             let formatted: Vec<Value> = results
diff --git a/crates/vestige-mcp/src/main.rs b/crates/vestige-mcp/src/main.rs
index e1329aa..e7e7697 100644
--- a/crates/vestige-mcp/src/main.rs
+++ b/crates/vestige-mcp/src/main.rs
@@ -243,6 +243,18 @@ async fn main() {
     let cognitive = Arc::new(Mutex::new(cognitive::CognitiveEngine::new()));
     info!("CognitiveEngine initialized (26 modules)");
 
+    // Load cross-encoder reranker in the background (downloads ~150MB on first run)
+    #[cfg(feature = "embeddings")]
+    {
+        let cog_clone = Arc::clone(&cognitive);
+        tokio::spawn(async move {
+            // Small delay so we don't block the stdio handshake
+            tokio::time::sleep(std::time::Duration::from_secs(1)).await;
+            let mut cog = cog_clone.lock().await;
+            cog.reranker.init_cross_encoder();
+        });
+    }
+
     // Create MCP server
     let server = McpServer::new(storage, cognitive);
 
diff --git a/crates/vestige-mcp/src/tools/search.rs b/crates/vestige-mcp/src/tools/search.rs
index ccbe27d..d72b264 100644
--- a/crates/vestige-mcp/src/tools/search.rs
+++ b/crates/vestige-mcp/src/tools/search.rs
@@ -162,8 +162,8 @@ pub async fn execute_hybrid(
         .hybrid_search(
             &args.query,
             args.limit.unwrap_or(10).clamp(1, 50),
-            args.keyword_weight.unwrap_or(0.5).clamp(0.0, 1.0),
-            args.semantic_weight.unwrap_or(0.5).clamp(0.0, 1.0),
+            args.keyword_weight.unwrap_or(0.3).clamp(0.0, 1.0),
+            args.semantic_weight.unwrap_or(0.7).clamp(0.0, 1.0),
         )
         .map_err(|e| e.to_string())?;
 
diff --git a/crates/vestige-mcp/src/tools/search_unified.rs b/crates/vestige-mcp/src/tools/search_unified.rs
index eba5423..b133b1f 100644
--- a/crates/vestige-mcp/src/tools/search_unified.rs
+++ b/crates/vestige-mcp/src/tools/search_unified.rs
@@ -127,9 +127,9 @@ pub async fn execute(
     let min_retention = args.min_retention.unwrap_or(0.0).clamp(0.0, 1.0);
     let min_similarity = args.min_similarity.unwrap_or(0.5).clamp(0.0, 1.0);
 
-    // Use balanced weights for hybrid search (keyword + semantic)
-    let keyword_weight = 0.5_f32;
-    let semantic_weight = 0.5_f32;
+    // Favor semantic search — research shows 0.3/0.7 outperforms equal weights
+    let keyword_weight = 0.3_f32;
+    let semantic_weight = 0.7_f32;
 
     // ====================================================================
     // STAGE 1: Hybrid search with 3x over-fetch for reranking pool
@@ -160,7 +160,7 @@ pub async fn execute(
     // ====================================================================
     // STAGE 2: Reranker (BM25-like rescoring, trim to requested limit)
     // ====================================================================
-    if let Ok(cog) = cognitive.try_lock() {
+    if let Ok(mut cog) = cognitive.try_lock() {
         let candidates: Vec<_> = filtered_results
             .iter()
             .map(|r| (r.clone(), r.node.content.clone()))
diff --git a/package.json b/package.json
index bd1ee40..bdd7fe1 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "vestige",
-  "version": "1.5.0",
+  "version": "1.6.0",
   "private": true,
   "description": "Cognitive memory for AI - MCP server with FSRS-6 spaced repetition",
   "author": "Sam Valladares",
diff --git a/packages/vestige-init/package.json b/packages/vestige-init/package.json
index 710ea45..05d3bc6 100644
--- a/packages/vestige-init/package.json
+++ b/packages/vestige-init/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@vestige/init",
-  "version": "1.5.0",
+  "version": "1.6.0",
   "description": "Give your AI a brain in 10 seconds — zero-config Vestige installer",
   "bin": {
     "vestige-init": "bin/init.js"
diff --git a/packages/vestige-mcp-npm/package.json b/packages/vestige-mcp-npm/package.json
index 415a0c8..418d97a 100644
--- a/packages/vestige-mcp-npm/package.json
+++ b/packages/vestige-mcp-npm/package.json
@@ -1,6 +1,6 @@
 {
   "name": "vestige-mcp-server",
-  "version": "1.5.0",
+  "version": "1.6.0",
   "description": "Vestige MCP Server - AI Memory System for Claude and other assistants",
   "bin": {
     "vestige-mcp": "bin/vestige-mcp.js",