diff --git a/packages/mnemopi/src/core/local-llm.ts b/packages/mnemopi/src/core/local-llm.ts
index 00817c49c..66fb39647 100644
--- a/packages/mnemopi/src/core/local-llm.ts
+++ b/packages/mnemopi/src/core/local-llm.ts
@@ -254,7 +254,7 @@ export function cleanOutput(text: string): string {
 		.replaceAll("<|user|>", "")
 		.replaceAll("</s>", "")
 		.trim()
-		.replace(/<think>[\s\S]*?<\/think>/gi, "")
+		.replace(/^(?:\s*<think>[\s\S]*?<\/think>)+\s*/i, "")
 		.replace(/^(Summarize the following memories.*?[.!?:]\s*)/is, "")
 		.replace(/^(Preserve facts.*?[.!?:]\s*)/is, "")
 		.replace(/^Source:.*?\n/im, "")
diff --git a/packages/mnemopi/test/local-llm.test.ts b/packages/mnemopi/test/local-llm.test.ts
index a8b80d2b9..e3f94135a 100644
--- a/packages/mnemopi/test/local-llm.test.ts
+++ b/packages/mnemopi/test/local-llm.test.ts
@@ -11,6 +11,7 @@ import {
 	callLocalLlm,
 	callRemoteLlm,
 	chunkMemoriesByBudget,
+	cleanOutput,
 	complete,
 	llmAvailable,
 	localGgufAvailable,
@@ -83,6 +84,10 @@ describe("local LLM TypeScript port", () => {
 		);
 	});
 
+	it("preserves a literal think tag that follows real content", () => {
+		expect(cleanOutput("The XML tag is <think>keep</think>")).toBe("The XML tag is <think>keep</think>");
+	});
+
 	it("keeps local GGUF unavailable and returns null for local completion", async () => {
 		expect(localGgufAvailable()).toBe(false);
 		expect(await callLocalLlm("prompt")).toBeNull();