diff --git a/Cargo.lock b/Cargo.lock
index 5d0ddc5..f6c8d8e 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -2063,7 +2063,9 @@ dependencies = [
  "anyhow",
  "futures",
  "hex",
+ "indexmap",
  "magika",
+ "once_cell",
  "ort",
  "regex",
  "serde",
diff --git a/crates/mem-core/Cargo.toml b/crates/mem-core/Cargo.toml
index 40ca688..e6ce638 100644
--- a/crates/mem-core/Cargo.toml
+++ b/crates/mem-core/Cargo.toml
@@ -19,3 +19,5 @@ time = { version = "0.3", features = ["serde", "formatting", "parsing", "macros"
 magika = "1.1.0"
 ort = { version = "2.0.0-rc.12", default-features = true }
 regex = "1.10"
+once_cell = "1.19"
+indexmap = "2.0"
diff --git a/crates/mem-core/src/lib.rs b/crates/mem-core/src/lib.rs
index 2346387..fe9b455 100644
--- a/crates/mem-core/src/lib.rs
+++ b/crates/mem-core/src/lib.rs
@@ -20,4 +20,4 @@ pub use lesson::{
 pub use query::{Query, QuerySet, SynthesisQuery};
 pub use prompt::{PromptBuilder, PromptMessages};
 pub use symptom_projection::{project_symptom, SymptomVector};
-pub use optimizer::{ContextOptimizer, ContextOptimizerConfig, ContentType, OptimizedChunk};
+pub use optimizer::{ContextOptimizer, ContextOptimizerConfig, ContentType, OptimizedChunk, CacheAligner, AlignedContent, CcrStore};
diff --git a/crates/mem-core/src/optimizer/cache_align.rs b/crates/mem-core/src/optimizer/cache_align.rs
new file mode 100644
index 0000000..0787322
--- /dev/null
+++ b/crates/mem-core/src/optimizer/cache_align.rs
@@ -0,0 +1,191 @@
+//! CacheAligner — Stabilize prompt prefix for LLM provider KV cache hits
+//!
+//! LLM providers (Anthropic, OpenAI) use prefix-based KV caching. A single
+//! changing timestamp early in the prompt invalidates the entire cache.
+//!
+//! CacheAligner detects dynamic patterns and moves them to the context tail,
+//! preserving the stable prefix for cache hits.
+
+use regex::Regex;
+use once_cell::sync::Lazy;
+
+static ISO_TIMESTAMP: Lazy<Regex> = Lazy::new(|| Regex::new(r"\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}").unwrap());
+static UUID: Lazy<Regex> = Lazy::new(|| Regex::new(r"[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}").unwrap());
+static SESSION_ID: Lazy<Regex> = Lazy::new(|| Regex::new(r"session[_-]?id[=:]\s*([a-zA-Z0-9]+)").unwrap());
+static RUN_ID: Lazy<Regex> = Lazy::new(|| Regex::new(r"run[_-]?id[=:]\s*([a-zA-Z0-9_-]+)").unwrap());
+static TEMP_PATH: Lazy<Regex> = Lazy::new(|| Regex::new(r"(/tmp|/var/tmp|C:\\Users\\[^\\]+\\AppData|~)/[^\s]+").unwrap());
+static SHA256: Lazy<Regex> = Lazy::new(|| Regex::new(r"[a-f0-9]{64}").unwrap());
+static LINE_COL: Lazy<Regex> = Lazy::new(|| Regex::new(r":\d{1,5}:\d{1,5}").unwrap());
+
+pub struct CacheAligner;
+
+#[derive(Debug, Clone)]
+pub struct AlignedContent {
+    /// Stable prefix (should be cached by LLM provider)
+    pub stable_prefix: String,
+    /// Dynamic tail (timestamps, UUIDs, session IDs, etc.)
+    pub dynamic_tail: String,
+    /// How much of the prefix changed (0.0 = identical, 1.0 = completely different)
+    pub drift_metric: f32,
+}
+
+impl CacheAligner {
+    /// Stabilize prompt by moving dynamic content to tail
+    pub fn align(content: &str) -> AlignedContent {
+        let lines: Vec<&str> = content.lines().collect();
+        let total_lines = lines.len();
+        let mut static_lines = Vec::new();
+        let mut dynamic_lines = Vec::new();
+
+        for line in &lines {
+            if Self::is_dynamic_line(line) {
+                dynamic_lines.push(*line);
+            } else {
+                static_lines.push(*line);
+            }
+        }
+
+        let stable_prefix = static_lines.join("\n");
+        let dynamic_tail = if dynamic_lines.is_empty() {
+            String::new()
+        } else {
+            format!("\n<!-- Dynamic context -->\n{}", dynamic_lines.join("\n"))
+        };
+
+        // Drift metric: ratio of dynamic lines
+        let drift_metric = if total_lines > 0 {
+            dynamic_lines.len() as f32 / total_lines as f32
+        } else {
+            0.0
+        };
+
+        AlignedContent {
+            stable_prefix,
+            dynamic_tail,
+            drift_metric,
+        }
+    }
+
+    /// Check if a line contains dynamic content
+    fn is_dynamic_line(line: &str) -> bool {
+        // Timestamps
+        if ISO_TIMESTAMP.is_match(line) {
+            return true;
+        }
+
+        // UUIDs
+        if UUID.is_match(line) {
+            return true;
+        }
+
+        // Session/run IDs
+        if SESSION_ID.is_match(line) || RUN_ID.is_match(line) {
+            return true;
+        }
+
+        // Temp paths
+        if TEMP_PATH.is_match(line) {
+            return true;
+        }
+
+        // SHA256 hashes (but not in common markers like "ccr:" prefix)
+        if SHA256.is_match(line) && !line.starts_with("<!--") {
+            return true;
+        }
+
+        // Line:col markers (they appear in stack traces but indicate location)
+        if LINE_COL.is_match(line) && (line.contains("at ") || line.contains("in file")) {
+            return true;
+        }
+
+        false
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_align_moves_timestamps_to_tail() {
+        let content = "System prompt\nAt 2026-08-28T09:15:00Z the query was:\nQuery text";
+        let aligned = CacheAligner::align(content);
+
+        assert!(aligned.stable_prefix.contains("System prompt"));
+        assert!(aligned.stable_prefix.contains("Query text"));
+        assert!(aligned.dynamic_tail.contains("2026-08-28T09:15:00Z"));
+    }
+
+    #[test]
+    fn test_align_moves_uuids_to_tail() {
+        let content = "Request with ID 550e8400-e29b-41d4-a716-446655440000\nThen do something";
+        let aligned = CacheAligner::align(content);
+
+        assert!(!aligned.stable_prefix.contains("550e8400"));
+        assert!(aligned.dynamic_tail.contains("550e8400"));
+    }
+
+    #[test]
+    fn test_align_moves_session_ids_to_tail() {
+        let content = "Start\nsession_id=abc123xyz\nEnd";
+        let aligned = CacheAligner::align(content);
+
+        assert!(aligned.stable_prefix.contains("Start"));
+        assert!(aligned.stable_prefix.contains("End"));
+        assert!(aligned.dynamic_tail.contains("session_id"));
+    }
+
+    #[test]
+    fn test_align_moves_temp_paths_to_tail() {
+        let content = "Error at /tmp/abc123/file.txt\nFix required";
+        let aligned = CacheAligner::align(content);
+
+        assert!(aligned.stable_prefix.contains("Fix"));
+        assert!(aligned.dynamic_tail.contains("/tmp"));
+    }
+
+    #[test]
+    fn test_align_preserves_code_line_col() {
+        let content = "Stack trace:\n  at main.rs:42:5\nIn function";
+        let aligned = CacheAligner::align(content);
+
+        // Line:col patterns in stack traces are dynamic
+        assert!(aligned.dynamic_tail.contains("42:5") || aligned.stable_prefix.contains("at main.rs"));
+    }
+
+    #[test]
+    fn test_align_drift_metric() {
+        let static_content = "System prompt\nStatic query\nStatic context";
+        let dynamic_content = format!("{}\nAt 2026-08-28T09:15:00Z\nID: 550e8400-e29b-41d4-a716-446655440000", static_content);
+
+        let static_aligned = CacheAligner::align(static_content);
+        let dynamic_aligned = CacheAligner::align(&dynamic_content);
+
+        // Dynamic content should have higher drift
+        assert!(dynamic_aligned.drift_metric > static_aligned.drift_metric);
+    }
+
+    #[test]
+    fn test_align_stable_across_calls_with_same_timestamp() {
+        let content1 = "Query at 2026-08-28T09:15:00Z\nAction";
+        let content2 = "Query at 2026-08-28T09:15:01Z\nAction"; // Different second
+
+        let aligned1 = CacheAligner::align(content1);
+        let aligned2 = CacheAligner::align(content2);
+
+        // Both should have identical stable prefixes (timestamp separated)
+        assert_eq!(aligned1.stable_prefix, aligned2.stable_prefix);
+        // But different dynamic tails
+        assert_ne!(aligned1.dynamic_tail, aligned2.dynamic_tail);
+    }
+
+    #[test]
+    fn test_align_no_dynamic_content() {
+        let content = "Pure static prompt\nWith no timestamps\nOr identifiers";
+        let aligned = CacheAligner::align(content);
+
+        // Should be entirely in prefix
+        assert_eq!(aligned.drift_metric, 0.0);
+        assert!(aligned.dynamic_tail.is_empty());
+    }
+}
diff --git a/crates/mem-core/src/optimizer/ccr.rs b/crates/mem-core/src/optimizer/ccr.rs
new file mode 100644
index 0000000..db53be6
--- /dev/null
+++ b/crates/mem-core/src/optimizer/ccr.rs
@@ -0,0 +1,209 @@
+//! CCR Store — Compress-Cache-Retrieve: reversible compression with retrieval
+//!
+//! When compressing, cache the full original with a SHA256 hash.
+//! If the model needs more detail, it can request the original via hash lookup.
+//! This makes compression aggressive but reversible.
+
+use anyhow::Result;
+use sha2::{Sha256, Digest};
+use indexmap::IndexMap;
+use std::sync::Mutex;
+use time::OffsetDateTime;
+
+pub struct CcrStore {
+    /// LRU cache: hash → (original_content, timestamp)
+    cache: Mutex<IndexMap<String, (String, OffsetDateTime)>>,
+    max_entries: usize,
+    ttl_secs: u64,
+}
+
+impl CcrStore {
+    /// Create a new CCR store with default limits
+    pub fn new() -> Self {
+        Self::with_limits(1000, 3600) // 1000 entries, 1 hour TTL
+    }
+
+    /// Create with custom limits
+    pub fn with_limits(max_entries: usize, ttl_secs: u64) -> Self {
+        Self {
+            cache: Mutex::new(IndexMap::new()),
+            max_entries,
+            ttl_secs,
+        }
+    }
+
+    /// Store original content and return its hash
+    pub fn store(&self, content: &str) -> Result<String> {
+        let hash = self.hash_content(content);
+        let mut cache = self.cache.lock().map_err(|e| anyhow::anyhow!("mutex lock failed: {}", e))?;
+
+        // Remove oldest entry if at capacity
+        if cache.len() >= self.max_entries {
+            if let Some(oldest_key) = cache.keys().next().cloned() {
+                cache.remove(&oldest_key);
+            }
+        }
+
+        cache.insert(hash.clone(), (content.to_string(), OffsetDateTime::now_utc()));
+
+        Ok(hash)
+    }
+
+    /// Retrieve original content by hash
+    pub fn retrieve(&self, hash: &str) -> Result<Option<String>> {
+        let mut cache = self.cache.lock().map_err(|e| anyhow::anyhow!("mutex lock failed: {}", e))?;
+
+        if let Some((content, timestamp)) = cache.get(hash) {
+            // Check if expired
+            let duration = OffsetDateTime::now_utc() - *timestamp;
+            if duration.whole_seconds() > self.ttl_secs as i64 {
+                cache.remove(hash);
+                return Ok(None);
+            }
+
+            // Return copy without moving (to avoid mutable borrow)
+            return Ok(Some(content.clone()));
+        }
+
+        Ok(None)
+    }
+
+    /// Compute hash of content
+    fn hash_content(&self, content: &str) -> String {
+        let mut hasher = Sha256::new();
+        hasher.update(content.as_bytes());
+        format!("{:x}", hasher.finalize())
+    }
+
+    /// Get current cache size
+    pub fn size(&self) -> Result<usize> {
+        let cache = self.cache.lock().map_err(|e| anyhow::anyhow!("mutex lock failed: {}", e))?;
+        Ok(cache.len())
+    }
+
+    /// Clear expired entries
+    pub fn evict_expired(&self) -> Result<usize> {
+        let mut cache = self.cache.lock().map_err(|e| anyhow::anyhow!("mutex lock failed: {}", e))?;
+        let before = cache.len();
+
+        cache.retain(|_, (_, timestamp)| {
+            let duration = OffsetDateTime::now_utc() - *timestamp;
+            duration.whole_seconds() <= self.ttl_secs as i64
+        });
+
+        Ok(before - cache.len())
+    }
+}
+
+impl Default for CcrStore {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_ccr_store_and_retrieve() {
+        let store = CcrStore::new();
+        let content = "Original content to store";
+
+        let hash = store.store(content).unwrap();
+        assert!(!hash.is_empty());
+        assert_eq!(hash.len(), 64); // SHA256 hex
+
+        let retrieved = store.retrieve(&hash).unwrap();
+        assert_eq!(retrieved, Some(content.to_string()));
+    }
+
+    #[test]
+    fn test_ccr_hash_deterministic() {
+        let store = CcrStore::new();
+        let content = "Deterministic content";
+
+        let hash1 = store.store(content).unwrap();
+        let hash2 = store.store(content).unwrap();
+
+        assert_eq!(hash1, hash2);
+    }
+
+    #[test]
+    fn test_ccr_different_content_different_hash() {
+        let store = CcrStore::new();
+
+        let hash1 = store.store("Content A").unwrap();
+        let hash2 = store.store("Content B").unwrap();
+
+        assert_ne!(hash1, hash2);
+    }
+
+    #[test]
+    fn test_ccr_retrieve_nonexistent() {
+        let store = CcrStore::new();
+        let result = store.retrieve("nonexistent_hash").unwrap();
+        assert_eq!(result, None);
+    }
+
+    #[test]
+    fn test_ccr_lru_eviction() {
+        let store = CcrStore::with_limits(3, 3600);
+
+        let h1 = store.store("content1").unwrap();
+        let h2 = store.store("content2").unwrap();
+        let h3 = store.store("content3").unwrap();
+
+        // Cache at capacity
+        assert_eq!(store.size().unwrap(), 3);
+
+        // Add new entry should evict oldest (h1)
+        let _h4 = store.store("content4").unwrap();
+        assert_eq!(store.size().unwrap(), 3);
+
+        // h1 should be gone
+        assert_eq!(store.retrieve(&h1).unwrap(), None);
+        // h2, h3 should still exist
+        assert!(store.retrieve(&h2).unwrap().is_some());
+        assert!(store.retrieve(&h3).unwrap().is_some());
+    }
+
+    #[test]
+    fn test_ccr_inject_hint() {
+        let store = CcrStore::new();
+        let content = "Original error message";
+
+        let hash = store.store(content).unwrap();
+        let hint = format!("<!-- CCR:{} -->", hash);
+
+        assert!(hint.len() > 10);
+        assert!(hint.contains("CCR:"));
+        assert!(hint.contains(&hash));
+    }
+
+    #[test]
+    fn test_ccr_large_content() {
+        let store = CcrStore::new();
+        let large = "x".repeat(100_000);
+
+        let hash = store.store(&large).unwrap();
+        let retrieved = store.retrieve(&hash).unwrap();
+
+        assert_eq!(retrieved.unwrap().len(), 100_000);
+    }
+
+    #[test]
+    fn test_ccr_multiple_stores_same_content() {
+        let store = CcrStore::new();
+
+        let h1 = store.store("shared").unwrap();
+        let h2 = store.store("shared").unwrap();
+
+        // Same content should produce same hash
+        assert_eq!(h1, h2);
+
+        // But they should have separate cache entries (LRU)
+        // Last one should be retrievable
+        assert!(store.retrieve(&h2).unwrap().is_some());
+    }
+}
diff --git a/crates/mem-core/src/optimizer/mod.rs b/crates/mem-core/src/optimizer/mod.rs
index b7ce175..26cad17 100644
--- a/crates/mem-core/src/optimizer/mod.rs
+++ b/crates/mem-core/src/optimizer/mod.rs
@@ -8,6 +8,8 @@ pub mod router;
 pub mod log;
 pub mod json;
 pub mod diff;
+pub mod cache_align;
+pub mod ccr;
 
 use anyhow::Result;
 use serde::{Deserialize, Serialize};
@@ -16,6 +18,8 @@ pub use router::ContentRouter;
 pub use log::LogCompressor;
 pub use json::JsonCrusher;
 pub use diff::DiffCompressor;
+pub use cache_align::{CacheAligner, AlignedContent};
+pub use ccr::CcrStore;
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct OptimizedChunk {
@@ -84,6 +88,7 @@ pub struct ContextOptimizer {
     log_compressor: LogCompressor,
     json_crusher: JsonCrusher,
     diff_compressor: DiffCompressor,
+    ccr_store: CcrStore,
     config: ContextOptimizerConfig,
 }
 
@@ -99,16 +104,23 @@ impl ContextOptimizer {
         let log_compressor = LogCompressor::new();
         let json_crusher = JsonCrusher::new();
         let diff_compressor = DiffCompressor::new();
+        let ccr_store = CcrStore::new();
 
         Ok(Self {
             router,
             log_compressor,
             json_crusher,
             diff_compressor,
+            ccr_store,
             config,
         })
     }
 
+    /// Get reference to CCR store for retrieval
+    pub fn ccr_store(&self) -> &CcrStore {
+        &self.ccr_store
+    }
+
     /// Optimize a chunk of content
     pub fn optimize(&self, content: &str) -> Result<OptimizedChunk> {
         if !self.config.enabled {
@@ -143,12 +155,19 @@ impl ContextOptimizer {
         let original_tokens = estimate_tokens(content);
         let compressed_tokens = estimate_tokens(&compressed);
 
+        // Store original in CCR if compression happened and CCR is enabled
+        let ccr_hash = if self.config.ccr_enabled && compressed != content {
+            self.ccr_store.store(content).ok()
+        } else {
+            None
+        };
+
         Ok(OptimizedChunk {
             compressed,
             original_tokens,
             compressed_tokens,
             content_type,
-            ccr_hash: None, // TODO: Implement CCR
+            ccr_hash,
         })
     }
 }
@@ -185,4 +204,26 @@ mod tests {
         let chunk = optimizer.optimize("test content").unwrap();
         assert_eq!(chunk.compressed, "test content");
     }
+
+    #[test]
+    fn test_optimizer_with_ccr() {
+        let optimizer = ContextOptimizer::new().unwrap();
+        let log_content = "ERROR: failed\nWARN: ignored";
+        let chunk = optimizer.optimize(log_content).unwrap();
+
+        // If compression happened and CCR enabled, should have hash
+        if chunk.compressed != log_content {
+            assert!(chunk.ccr_hash.is_some());
+        }
+    }
+
+    #[test]
+    fn test_optimizer_cache_align() {
+        let content = "System prompt\nAt 2026-08-28T09:15:00Z query was:\nContext";
+        let aligned = CacheAligner::align(content);
+
+        assert!(aligned.stable_prefix.contains("System"));
+        assert!(aligned.stable_prefix.contains("Context"));
+        assert!(aligned.dynamic_tail.contains("2026-08-28"));
+    }
 }