diff --git a/crates/tui/src/commands/session.rs b/crates/tui/src/commands/session.rs
index 098b00eb..306f76cf 100644
--- a/crates/tui/src/commands/session.rs
+++ b/crates/tui/src/commands/session.rs
@@ -226,7 +226,7 @@ pub fn load(app: &mut App, path: Option<&str>) -> CommandResult {
         }
     };
 
-    app.api_messages.clone_from(&session.messages);
+    app.api_messages = session.messages.clone().into();
     app.clear_history();
     let cells_to_add: Vec<_> = app
         .api_messages
diff --git a/crates/tui/src/core/engine.rs b/crates/tui/src/core/engine.rs
index 6f89fe19..ba9a2f88 100644
--- a/crates/tui/src/core/engine.rs
+++ b/crates/tui/src/core/engine.rs
@@ -1302,7 +1302,7 @@ impl Engine {
                     } else if messages.is_empty() && system_prompt.is_none() {
                         self.session.id = uuid::Uuid::new_v4().to_string();
                     }
-                    self.session.messages = messages;
+                    self.session.messages = messages.into();
                     self.session.compaction_summary_prompt =
                         extract_compaction_summary_prompt(system_prompt.clone());
                     self.session.system_prompt = system_prompt;
@@ -1351,7 +1351,7 @@ impl Engine {
                         }
                     }
                     if let Some(idx) = cut {
-                        self.session.messages.truncate(idx);
+                        self.session.messages.truncate_to(idx);
                         self.session.bump_messages_revision();
                     }
                     // Now dispatch the new message as a normal send,
@@ -1398,7 +1398,7 @@ impl Engine {
             .tx_event
             .send(Event::SessionUpdated {
                 session_id: self.session.id.clone(),
-                messages: self.session.messages.clone(),
+                messages: self.session.messages.clone().into(),
                 system_prompt: self.session.system_prompt.clone(),
                 model: self.session.model.clone(),
                 workspace: self.session.workspace.clone(),
@@ -1938,7 +1938,7 @@ impl Engine {
             Ok(result) => {
                 if !result.messages.is_empty() || self.session.messages.is_empty() {
                     let messages_after = result.messages.len();
-                    self.session.messages = result.messages;
+                    self.session.messages = result.messages.into();
                     self.merge_compaction_summary(result.summary_prompt);
                     self.emit_session_updated().await;
                     let removed = messages_before.saturating_sub(messages_after);
@@ -2034,7 +2034,7 @@ impl Engine {
         {
             Ok(result) => {
                 let messages_after = result.messages.len();
-                self.session.messages = result.messages;
+                self.session.messages = result.messages.into();
                 self.emit_session_updated().await;
 
                 let summary = format!(
@@ -2088,7 +2088,7 @@ impl Engine {
         while self.session.messages.len() > MIN_RECENT_MESSAGES_TO_KEEP
             && self.estimated_input_tokens() > target_input_budget
         {
-            self.session.messages.remove(0);
+            self.session.messages.trim_front(1);
             self.session.bump_messages_revision();
             removed = removed.saturating_add(1);
         }
@@ -2110,7 +2110,7 @@ impl Engine {
 
         let mut retries_used = 0u32;
         let mut summary_prompt = None;
-        let mut compacted_messages = self.session.messages.clone();
+        let mut compacted_messages: Vec<Message> = self.session.messages.clone().into();
 
         let mut forced_config = self.config.compaction.clone();
         forced_config.enabled = true;
@@ -2145,7 +2145,7 @@ impl Engine {
         }
 
         if !compacted_messages.is_empty() || self.session.messages.is_empty() {
-            self.session.messages = compacted_messages;
+            self.session.messages = compacted_messages.into();
         }
         self.merge_compaction_summary(summary_prompt);
 
@@ -2219,7 +2219,7 @@ impl Engine {
             self.session.model.clone(),
             self.session.workspace.clone(),
             self.session.system_prompt.clone(),
-            self.session.messages.clone(),
+            self.session.messages.clone().into(),
         ))
         .with_cancel_token(self.cancel_token.clone())
         .with_trusted_external_paths(trusted_external_paths);
diff --git a/crates/tui/src/core/engine/capacity_flow.rs b/crates/tui/src/core/engine/capacity_flow.rs
index 514ad120..3385f1e7 100644
--- a/crates/tui/src/core/engine/capacity_flow.rs
+++ b/crates/tui/src/core/engine/capacity_flow.rs
@@ -415,7 +415,7 @@ impl Engine {
             {
                 Ok(result) => {
                     if !result.messages.is_empty() || self.session.messages.is_empty() {
-                        self.session.messages = result.messages;
+                        self.session.messages = result.messages.into();
                         self.merge_compaction_summary(result.summary_prompt);
                         refreshed = true;
                     }
diff --git a/crates/tui/src/core/engine/tests.rs b/crates/tui/src/core/engine/tests.rs
index 212042c4..95e752be 100644
--- a/crates/tui/src/core/engine/tests.rs
+++ b/crates/tui/src/core/engine/tests.rs
@@ -1799,12 +1799,13 @@ fn runtime_prompt_is_projected_without_persisting_to_session_messages() {
             text: "summary after compaction".to_string(),
             cache_control: None,
         }],
-    }];
+    }]
+    .into();
     let stored = engine.session.messages.clone();
 
     let request_messages = engine.messages_with_turn_metadata();
 
-    assert_eq!(engine.session.messages, stored);
+    assert_eq!(&*engine.session.messages, &*stored);
     assert_eq!(request_messages.len(), stored.len() + 1);
     assert!(
         request_messages
@@ -2488,7 +2489,7 @@ fn messages_with_turn_metadata_preserves_stored_messages_for_prefix_cache() {
     let first_request = engine.messages_with_turn_metadata();
     assert_eq!(
         &first_request[..engine.session.messages.len()],
-        engine.session.messages.as_slice()
+        &engine.session.messages[..]
     );
     assert_eq!(first_request.len(), engine.session.messages.len() + 1);
     assert_eq!(first_request.first(), Some(&first_user));
@@ -2514,7 +2515,7 @@ fn messages_with_turn_metadata_preserves_stored_messages_for_prefix_cache() {
     let second_request = engine.messages_with_turn_metadata();
     assert_eq!(
         &second_request[..engine.session.messages.len()],
-        engine.session.messages.as_slice()
+        &engine.session.messages[..]
     );
     assert_eq!(second_request.len(), engine.session.messages.len() + 1);
     assert_eq!(second_request.first(), Some(&first_user));
diff --git a/crates/tui/src/core/engine/turn_loop.rs b/crates/tui/src/core/engine/turn_loop.rs
index 53cdf927..afb1a9b0 100644
--- a/crates/tui/src/core/engine/turn_loop.rs
+++ b/crates/tui/src/core/engine/turn_loop.rs
@@ -153,7 +153,7 @@ impl Engine {
                         // Only update if we got valid messages (never corrupt state)
                         if !result.messages.is_empty() || self.session.messages.is_empty() {
                             let auto_messages_after = result.messages.len();
-                            self.session.messages = result.messages;
+                            self.session.messages = result.messages.into();
                             self.merge_compaction_summary(result.summary_prompt);
                             self.emit_session_updated().await;
                             let removed = auto_messages_before.saturating_sub(auto_messages_after);
@@ -2312,7 +2312,7 @@ impl Engine {
         // messages. This preserves the stable prefix through all stored
         // messages while avoiding strict chat templates that only allow
         // system messages at messages[0].
-        let mut messages = self.session.messages.clone();
+        let mut messages: Vec<Message> = self.session.messages.clone().into();
         messages.push(self.runtime_prompt_message());
         messages
     }
diff --git a/crates/tui/src/core/session.rs b/crates/tui/src/core/session.rs
index 67ffb7ad..a0cd4d94 100644
--- a/crates/tui/src/core/session.rs
+++ b/crates/tui/src/core/session.rs
@@ -5,7 +5,7 @@
 use crate::models::{Message, SystemPrompt, Usage};
 use crate::prefix_cache::PrefixStabilityManager;
 use crate::project_context::{ProjectContext, load_project_context_with_parents};
-use crate::prompt_zones::FrozenPrefix;
+use crate::prompt_zones::{AppendLog, FrozenPrefix};
 use crate::tui::approval::ApprovalMode;
 use crate::working_set::WorkingSet;
 use std::path::PathBuf;
@@ -40,8 +40,8 @@ pub struct Session {
     /// Persisted summary blocks generated by context compaction.
     pub compaction_summary_prompt: Option<SystemPrompt>,
 
-    /// Conversation history (API format)
-    pub messages: Vec<Message>,
+    /// Conversation history (API format), backed by AppendLog (#2264).
+    pub messages: AppendLog,
 
     /// Total tokens used in this session
     pub total_usage: SessionUsage,
@@ -145,7 +145,7 @@ impl Session {
             system_prompt: None,
             system_prompt_override: false,
             compaction_summary_prompt: None,
-            messages: Vec::new(),
+            messages: AppendLog::new(),
             total_usage: SessionUsage::default(),
             allow_shell,
             trust_mode,
@@ -179,7 +179,7 @@ impl Session {
     /// invalidate atomically.
     #[allow(dead_code)]
     pub fn replace_messages(&mut self, messages: Vec<Message>) {
-        self.messages = messages;
+        self.messages = messages.into();
         self.messages_revision = self.messages_revision.saturating_add(1);
     }
 
diff --git a/crates/tui/src/prompt_zones.rs b/crates/tui/src/prompt_zones.rs
index 581ac355..e9c9c04a 100644
--- a/crates/tui/src/prompt_zones.rs
+++ b/crates/tui/src/prompt_zones.rs
@@ -195,16 +195,17 @@ impl std::fmt::Display for PrefixDrift {
 
 // ── AppendLog ──────────────────────────────────────────────────────────
 
-/// Append-only conversation history. Only exposes `push`-style mutations.
+/// Append-only conversation history. Derefs to `&[Message]` via
+/// [`Deref`](std::ops::Deref) for transparent read access; mutations go
+/// through explicit methods (`push`, `truncate_to`, `trim_front`, `clear`)
+/// whose names make cache impact obvious.
 ///
-/// **Phase 1 scaffolding** — not yet wired into the engine request path.
-#[allow(dead_code)]
+/// Phase 4: backing store for `Session.messages` (#2264).
 #[derive(Debug, Clone)]
 pub struct AppendLog {
     messages: Vec<Message>,
 }
 
-#[allow(dead_code)]
 impl AppendLog {
     pub fn new() -> Self {
         Self {
@@ -216,27 +217,53 @@ impl AppendLog {
         Self { messages }
     }
 
+    /// Append a message to the log. A single-message push is the cheapest
+    /// mutation for prefix-cache stability — it extends the byte sequence
+    /// without disturbing earlier turns.
     pub fn push(&mut self, message: Message) {
         self.messages.push(message);
     }
 
-    #[must_use]
-    pub fn len(&self) -> usize {
-        self.messages.len()
+    /// Append multiple messages in one operation (fewer cache-line
+    /// invalidations than repeated `push`).
+    pub fn push_batch(&mut self, batch: Vec<Message>) {
+        self.messages.extend(batch);
     }
 
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        self.messages.is_empty()
+    /// Truncate to keep only the most recent `new_len` messages.
+    /// Discards older messages (and their prefix-cache contribution)
+    /// from the front.
+    pub fn truncate_to(&mut self, new_len: usize) {
+        self.messages.truncate(new_len);
     }
 
-    pub fn iter(&self) -> impl Iterator<Item = &Message> {
-        self.messages.iter()
+    /// Remove `count` messages from the front (oldest first).
+    /// Cache-destroying: drops the prefix that earlier turns share.
+    pub fn trim_front(&mut self, count: usize) {
+        if count >= self.messages.len() {
+            self.messages.clear();
+        } else {
+            self.messages.drain(0..count);
+        }
     }
 
+    /// Remove all messages. Resets cache state completely.
+    pub fn clear(&mut self) {
+        self.messages.clear();
+    }
+
+    /// Return a mutable reference to the last message, if any.
+    /// Prefer this over `last_mut()` on the inner vec — the name signals
+    /// that only the most recent turn's content is being modified.
     #[must_use]
-    pub fn as_slice(&self) -> &[Message] {
-        &self.messages
+    pub fn last_mut(&mut self) -> Option<&mut Message> {
+        self.messages.last_mut()
+    }
+
+    /// Consume and return the inner `Vec<Message>`.
+    #[must_use]
+    pub fn into_inner(self) -> Vec<Message> {
+        self.messages
     }
 }
 
@@ -246,6 +273,26 @@ impl Default for AppendLog {
     }
 }
 
+impl From<Vec<Message>> for AppendLog {
+    fn from(messages: Vec<Message>) -> Self {
+        Self { messages }
+    }
+}
+
+impl From<AppendLog> for Vec<Message> {
+    fn from(log: AppendLog) -> Self {
+        log.messages
+    }
+}
+
+impl std::ops::Deref for AppendLog {
+    type Target = Vec<Message>;
+
+    fn deref(&self) -> &Self::Target {
+        &self.messages
+    }
+}
+
 // ── TurnScratch ────────────────────────────────────────────────────────
 
 /// Per-turn ephemeral data. Cleared at every turn boundary.