diff --git a/tests/agent_e2e.rs b/tests/agent_e2e.rs
index 9ca3287..6bdfb36 100644
--- a/tests/agent_e2e.rs
+++ b/tests/agent_e2e.rs
@@ -13,11 +13,15 @@ use serde_json::json;
 use std::sync::{Arc, Mutex};
 use zeroclaw::agent::agent::Agent;
 use zeroclaw::agent::dispatcher::{NativeToolDispatcher, XmlToolDispatcher};
+use zeroclaw::agent::memory_loader::MemoryLoader;
 use zeroclaw::config::MemoryConfig;
 use zeroclaw::memory;
 use zeroclaw::memory::Memory;
 use zeroclaw::observability::{NoopObserver, Observer};
-use zeroclaw::providers::{ChatRequest, ChatResponse, Provider, ToolCall};
+use zeroclaw::providers::traits::ChatMessage;
+use zeroclaw::providers::{
+    ChatRequest, ChatResponse, ConversationMessage, Provider, ProviderRuntimeOptions, ToolCall,
+};
 use zeroclaw::tools::{Tool, ToolResult};
 
 // ─────────────────────────────────────────────────────────────────────────────
@@ -138,6 +142,79 @@ impl Tool for CountingTool {
     }
 }
 
+/// Mock provider that returns scripted responses AND records every request.
+/// Pattern from `ScriptedProvider` in `src/agent/tests.rs`.
+struct RecordingProvider {
+    responses: Mutex<Vec<ChatResponse>>,
+    recorded_requests: Arc<Mutex<Vec<Vec<ChatMessage>>>>,
+}
+
+impl RecordingProvider {
+    fn new(responses: Vec<ChatResponse>) -> (Self, Arc<Mutex<Vec<Vec<ChatMessage>>>>) {
+        let recorded = Arc::new(Mutex::new(Vec::new()));
+        let provider = Self {
+            responses: Mutex::new(responses),
+            recorded_requests: recorded.clone(),
+        };
+        (provider, recorded)
+    }
+}
+
+#[async_trait]
+impl Provider for RecordingProvider {
+    async fn chat_with_system(
+        &self,
+        _system_prompt: Option<&str>,
+        _message: &str,
+        _model: &str,
+        _temperature: f64,
+    ) -> Result<String> {
+        Ok("fallback".into())
+    }
+
+    async fn chat(
+        &self,
+        request: ChatRequest<'_>,
+        _model: &str,
+        _temperature: f64,
+    ) -> Result<ChatResponse> {
+        self.recorded_requests
+            .lock()
+            .unwrap()
+            .push(request.messages.to_vec());
+
+        let mut guard = self.responses.lock().unwrap();
+        if guard.is_empty() {
+            return Ok(ChatResponse {
+                text: Some("done".into()),
+                tool_calls: vec![],
+            });
+        }
+        Ok(guard.remove(0))
+    }
+}
+
+/// Mock memory loader that returns a static context string,
+/// simulating RAG recall without a real memory backend.
+struct StaticMemoryLoader {
+    context: String,
+}
+
+impl StaticMemoryLoader {
+    fn new(context: &str) -> Self {
+        Self {
+            context: context.to_string(),
+        }
+    }
+}
+
+#[async_trait]
+impl MemoryLoader for StaticMemoryLoader {
+    async fn load_context(&self, _memory: &dyn Memory, _user_message: &str) -> Result<String> {
+        Ok(self.context.clone())
+    }
+}
+
 // ─────────────────────────────────────────────────────────────────────────────
 // Test helpers
 // ─────────────────────────────────────────────────────────────────────────────
@@ -192,6 +269,26 @@ fn build_agent_xml(provider: Box<dyn Provider>, tools: Vec<Box<dyn Tool>>) -> Ag
         .unwrap()
 }
 
+fn build_recording_agent(
+    provider: Box<dyn Provider>,
+    tools: Vec<Box<dyn Tool>>,
+    memory_loader: Option<Box<dyn MemoryLoader>>,
+) -> Agent {
+    let mut builder = Agent::builder()
+        .provider(provider)
+        .tools(tools)
+        .memory(make_memory())
+        .observer(make_observer())
+        .tool_dispatcher(Box::new(NativeToolDispatcher))
+        .workspace_dir(std::env::temp_dir());
+
+    if let Some(loader) = memory_loader {
+        builder = builder.memory_loader(loader);
+    }
+
+    builder.build().unwrap()
+}
+
 // ═════════════════════════════════════════════════════════════════════════════
 // E2E smoke tests — full agent turn cycle
 // ═════════════════════════════════════════════════════════════════════════════
@@ -352,3 +449,243 @@ async fn e2e_parallel_tool_dispatch() {
     );
     assert_eq!(*count.lock().unwrap(), 2);
 }
+
+// ═════════════════════════════════════════════════════════════════════════════
+// Multi-turn history fidelity & memory enrichment tests
+// ═════════════════════════════════════════════════════════════════════════════
+
+/// Validates that multi-turn conversation correctly accumulates history
+/// and passes growing message sequences to the provider on each turn.
+#[tokio::test]
+async fn e2e_multi_turn_history_fidelity() {
+    let (provider, recorded) = RecordingProvider::new(vec![
+        text_response("response 1"),
+        text_response("response 2"),
+        text_response("response 3"),
+    ]);
+
+    let mut agent = build_recording_agent(Box::new(provider), vec![], None);
+
+    let r1 = agent.turn("msg 1").await.unwrap();
+    assert_eq!(r1, "response 1");
+
+    let r2 = agent.turn("msg 2").await.unwrap();
+    assert_eq!(r2, "response 2");
+
+    let r3 = agent.turn("msg 3").await.unwrap();
+    assert_eq!(r3, "response 3");
+
+    let requests = recorded.lock().unwrap();
+    assert_eq!(requests.len(), 3, "Provider should receive 3 requests");
+
+    // Request 1: system + user("msg 1")
+    let req1 = &requests[0];
+    assert!(req1.len() >= 2);
+    assert_eq!(req1[0].role, "system");
+    assert_eq!(req1[1].role, "user");
+    assert!(req1[1].content.contains("msg 1"));
+
+    // Request 2: system + user("msg 1") + assistant("response 1") + user("msg 2")
+    let req2 = &requests[1];
+    let req2_users: Vec<&ChatMessage> = req2.iter().filter(|m| m.role == "user").collect();
+    let req2_assts: Vec<&ChatMessage> = req2.iter().filter(|m| m.role == "assistant").collect();
+    assert_eq!(req2_users.len(), 2, "Request 2: expected 2 user messages");
+    assert_eq!(
+        req2_assts.len(),
+        1,
+        "Request 2: expected 1 assistant message"
+    );
+    assert!(req2_users[0].content.contains("msg 1"));
+    assert!(req2_users[1].content.contains("msg 2"));
+    assert_eq!(req2_assts[0].content, "response 1");
+
+    // Request 3: full history — 3 user + 2 assistant messages
+    let req3 = &requests[2];
+    let req3_users: Vec<&ChatMessage> = req3.iter().filter(|m| m.role == "user").collect();
+    let req3_assts: Vec<&ChatMessage> = req3.iter().filter(|m| m.role == "assistant").collect();
+    assert_eq!(req3_users.len(), 3, "Request 3: expected 3 user messages");
+    assert_eq!(
+        req3_assts.len(),
+        2,
+        "Request 3: expected 2 assistant messages"
+    );
+    assert!(req3_users[0].content.contains("msg 1"));
+    assert!(req3_users[1].content.contains("msg 2"));
+    assert!(req3_users[2].content.contains("msg 3"));
+    assert_eq!(req3_assts[0].content, "response 1");
+    assert_eq!(req3_assts[1].content, "response 2");
+
+    // Verify agent history: system + 3*(user + assistant) = 7
+    let history = agent.history();
+    assert_eq!(history.len(), 7);
+    assert!(matches!(&history[0], ConversationMessage::Chat(c) if c.role == "system"));
+    assert!(matches!(&history[1], ConversationMessage::Chat(c) if c.role == "user"));
+    assert!(matches!(&history[2], ConversationMessage::Chat(c) if c.role == "assistant"));
+    assert!(
+        matches!(&history[6], ConversationMessage::Chat(c) if c.role == "assistant" && c.content == "response 3")
+    );
+}
+
+/// Validates that a custom MemoryLoader injects RAG context into user
+/// messages before they reach the provider.
+#[tokio::test]
+async fn e2e_memory_enrichment_injects_context() {
+    let (provider, recorded) = RecordingProvider::new(vec![text_response("enriched response")]);
+
+    let memory_context = "[Memory context]\n- user_name: test_user\n\n";
+    let loader = StaticMemoryLoader::new(memory_context);
+
+    let mut agent = build_recording_agent(Box::new(provider), vec![], Some(Box::new(loader)));
+
+    let response = agent.turn("hello").await.unwrap();
+    assert_eq!(response, "enriched response");
+
+    // Provider received enriched message
+    let requests = recorded.lock().unwrap();
+    assert_eq!(requests.len(), 1);
+    let user_msg = requests[0].iter().find(|m| m.role == "user").unwrap();
+    assert!(
+        user_msg.content.starts_with("[Memory context]"),
+        "User message should start with memory context, got: {}",
+        user_msg.content,
+    );
+    assert!(
+        user_msg.content.contains("user_name: test_user"),
+        "User message should contain memory key-value pair",
+    );
+    assert!(
+        user_msg.content.ends_with("hello"),
+        "User message should end with original text, got: {}",
+        user_msg.content,
+    );
+
+    // Agent history also stores enriched message
+    let history = agent.history();
+    match &history[1] {
+        ConversationMessage::Chat(c) => {
+            assert_eq!(c.role, "user");
+            assert!(c.content.starts_with("[Memory context]"));
+            assert!(c.content.ends_with("hello"));
+        }
+        other => panic!("Expected Chat variant for user message, got: {other:?}"),
+    }
+}
+
+/// Validates multi-turn conversation with memory enrichment: every user
+/// message is enriched, and the provider sees the full enriched history.
+#[tokio::test]
+async fn e2e_multi_turn_with_memory_enrichment() {
+    let (provider, recorded) =
+        RecordingProvider::new(vec![text_response("answer 1"), text_response("answer 2")]);
+
+    let memory_context = "[Memory context]\n- project: zeroclaw\n\n";
+    let loader = StaticMemoryLoader::new(memory_context);
+
+    let mut agent = build_recording_agent(Box::new(provider), vec![], Some(Box::new(loader)));
+
+    let r1 = agent.turn("first question").await.unwrap();
+    assert_eq!(r1, "answer 1");
+
+    let r2 = agent.turn("second question").await.unwrap();
+    assert_eq!(r2, "answer 2");
+
+    let requests = recorded.lock().unwrap();
+    assert_eq!(requests.len(), 2);
+
+    // Turn 1: user message is enriched
+    let req1_user = requests[0].iter().find(|m| m.role == "user").unwrap();
+    assert!(req1_user.content.contains("[Memory context]"));
+    assert!(req1_user.content.contains("project: zeroclaw"));
+    assert!(req1_user.content.ends_with("first question"));
+
+    // Turn 2: both user messages enriched, assistant from turn 1 present
+    let req2_users: Vec<&ChatMessage> = requests[1].iter().filter(|m| m.role == "user").collect();
+    assert_eq!(req2_users.len(), 2, "Request 2 should have 2 user messages");
+
+    // Turn 1 user message still enriched in history
+    assert!(req2_users[0].content.contains("[Memory context]"));
+    assert!(req2_users[0].content.ends_with("first question"));
+
+    // Turn 2 user message also enriched
+    assert!(req2_users[1].content.contains("[Memory context]"));
+    assert!(req2_users[1].content.ends_with("second question"));
+
+    // Assistant response from turn 1 preserved
+    let req2_assts: Vec<&ChatMessage> = requests[1]
+        .iter()
+        .filter(|m| m.role == "assistant")
+        .collect();
+    assert_eq!(req2_assts.len(), 1);
+    assert_eq!(req2_assts[0].content, "answer 1");
+
+    // History: system + 2*(enriched_user + assistant) = 5
+    assert_eq!(agent.history().len(), 5);
+}
+
+/// Validates that empty memory context passes user message through unmodified.
+#[tokio::test]
+async fn e2e_empty_memory_context_passthrough() {
+    let (provider, recorded) = RecordingProvider::new(vec![text_response("plain response")]);
+
+    let loader = StaticMemoryLoader::new("");
+
+    let mut agent = build_recording_agent(Box::new(provider), vec![], Some(Box::new(loader)));
+
+    let response = agent.turn("hello").await.unwrap();
+    assert_eq!(response, "plain response");
+
+    let requests = recorded.lock().unwrap();
+    let user_msg = requests[0].iter().find(|m| m.role == "user").unwrap();
+    assert_eq!(
+        user_msg.content, "hello",
+        "Empty context should not prepend anything to user message",
+    );
+}
+
+// ═════════════════════════════════════════════════════════════════════════════
+// Live integration test — real OpenAI Codex API (requires credentials)
+// ═════════════════════════════════════════════════════════════════════════════
+
+/// Sends a real multi-turn conversation to OpenAI Codex and verifies
+/// the model retains context from earlier messages.
+///
+/// Requires valid OAuth credentials in `~/.zeroclaw/`.
+/// Run manually: `cargo test e2e_live_openai_codex_multi_turn -- --ignored`
+#[tokio::test]
+#[ignore]
+async fn e2e_live_openai_codex_multi_turn() {
+    use zeroclaw::providers::openai_codex::OpenAiCodexProvider;
+    use zeroclaw::providers::traits::Provider;
+
+    let provider = OpenAiCodexProvider::new(&ProviderRuntimeOptions::default());
+    let model = "gpt-5.3-codex";
+
+    // Turn 1: establish a fact
+    let messages_turn1 = vec![
+        ChatMessage::system("You are a concise assistant. Reply in one short sentence."),
+        ChatMessage::user("The secret word is \"zephyr\". Just confirm you noted it."),
+    ];
+    let response1 = provider
+        .chat_with_history(&messages_turn1, model, 0.0)
+        .await;
+    assert!(response1.is_ok(), "Turn 1 failed: {:?}", response1.err());
+    let r1 = response1.unwrap();
+    assert!(!r1.is_empty(), "Turn 1 returned empty response");
+
+    // Turn 2: ask the model to recall the fact
+    let messages_turn2 = vec![
+        ChatMessage::system("You are a concise assistant. Reply in one short sentence."),
+        ChatMessage::user("The secret word is \"zephyr\". Just confirm you noted it."),
+        ChatMessage::assistant(&r1),
+        ChatMessage::user("What is the secret word?"),
+    ];
+    let response2 = provider
+        .chat_with_history(&messages_turn2, model, 0.0)
+        .await;
+    assert!(response2.is_ok(), "Turn 2 failed: {:?}", response2.err());
+    let r2 = response2.unwrap().to_lowercase();
+    assert!(
+        r2.contains("zephyr"),
+        "Model should recall 'zephyr' from history, got: {r2}",
+    );
+}