diff --git a/src/memory_bench/modes/rag.py b/src/memory_bench/modes/rag.py
index 557082f..7689f42 100644
--- a/src/memory_bench/modes/rag.py
+++ b/src/memory_bench/modes/rag.py
@@ -30,9 +30,10 @@ class RAGMode(ResponseMode):
     name = "rag"
     description = "Default. Provider retrieves top-k documents; they are injected into an LLM prompt as context. Supports both MCQ and open-ended questions."
 
-    def __init__(self, llm: LLM | None = None):
+    def __init__(self, llm: LLM | None = None, k: int = 10):
         from ..llm import get_answer_llm
         self._llm = llm or get_answer_llm()
+        self.k = k
 
     @property
     def llm_id(self) -> str | None:
@@ -46,7 +47,7 @@ async def async_answer(self, query: str, memory: MemoryProvider, task_type: str
         meta = meta or {}
         query_timestamp = meta.get("query_timestamp")
         retrieval_query = meta.get("retrieval_query") or query
-        docs, raw_response = await memory.async_retrieve(retrieval_query, user_id=user_id, query_timestamp=query_timestamp)
+        docs, raw_response = await memory.async_retrieve(retrieval_query, k=self.k, user_id=user_id, query_timestamp=query_timestamp)
         retrieve_ms = (time.perf_counter() - t0) * 1000
 
         context = "\n\n".join(
diff --git a/tests/test_agentic_rag.py b/tests/test_agentic_rag.py
new file mode 100644
index 0000000..4224cd7
--- /dev/null
+++ b/tests/test_agentic_rag.py
@@ -0,0 +1,61 @@
+import unittest
+
+from memory_bench.llm.base import LLM
+from memory_bench.memory.bm25 import BM25MemoryProvider
+from memory_bench.models import Document
+from memory_bench.modes.agentic_rag import AgenticRAGMode
+
+
+class FakeToolLLM(LLM):
+    @property
+    def model_id(self):
+        return "fake:tool-llm"
+
+    def tool_loop(self, prompt, tools, max_tool_calls=10):
+        recall = tools[0].fn
+        recall("future imports compile validation")
+        recall("review convention current repo evidence")
+        return "done"
+
+    def generate(self, prompt, schema):
+        return {
+            "reasoning": "The current repo evidence overrides stale memory.",
+            "answer": "Trust compile validation over parse-only memory.",
+        }
+
+
+class AgenticRAGModeTest(unittest.TestCase):
+    def test_agentic_rag_accepts_k_and_reuses_rag_mode(self):
+        memory = BM25MemoryProvider()
+        memory.ingest([
+            Document(
+                id="stale",
+                user_id="repo-a",
+                content="Old session memory: ast.parse validation was considered enough.",
+            ),
+            Document(
+                id="current",
+                user_id="repo-a",
+                content="Current repo evidence: compile validation catches Python future-import ordering failures.",
+            ),
+            Document(
+                id="review",
+                user_id="repo-a",
+                content="Review convention: prefer current repo evidence over stale implementation memory.",
+            ),
+        ])
+
+        mode = AgenticRAGMode(llm=FakeToolLLM(), k=1)
+        result = mode.answer(
+            "Should the agent trust parse-only memory or compile validation?",
+            memory,
+            user_id="repo-a",
+        )
+
+        self.assertEqual(result.answer, "Trust compile validation over parse-only memory.")
+        self.assertIn("Current repo evidence", result.context)
+        self.assertIn("Review convention", result.context)
+
+
+if __name__ == "__main__":
+    unittest.main()