diff --git a/src/core/prompts/l1-dedup.ts b/src/core/prompts/l1-dedup.ts
index a04bda30..bd921948 100644
--- a/src/core/prompts/l1-dedup.ts
+++ b/src/core/prompts/l1-dedup.ts
@@ -65,7 +65,7 @@ export const CONFLICT_DETECTION_SYSTEM_PROMPT = `你是记忆冲突检测器。
 - target_ids：要删除替换的旧记忆 ID **数组**（可以 1 条或多条）。store/skip 时省略或为空。
 - merged_content：merge/update 时的最终记忆文本。store/skip 时省略。
 - merged_type：merge/update 后记忆应归属的 type。根据合并后内容本质判断。
-- merged_priority：merge/update 后的新优先级（0-100 整数，merge/update 时必填）。合并后信息更完整、更确定，通常应**酌情提升** priority（例如两条 priority 70 的记忆合并后可提升到 80）。参考标准：80-100（核心特质/重要事件），60-79（一般偏好/普通活动），<60（次要信息）。
+- merged_priority：merge/update 后的新优先级（0-100 整数，merge/update 时必填）。不要因为合并本身提高优先级；merged_priority 必须小于或等于参与合并/更新的新记忆与 target_ids 对应旧记忆中的最高 priority。参考标准：80-100（核心特质/重要事件），60-79（一般偏好/普通活动），<60（次要信息）。
 - merged_timestamps：合并后的时间戳数组。收集新记忆 + 所有被合并旧记忆的时间戳，去重排序。`;
 
 // ============================
diff --git a/src/core/record/l1-dedup.test.ts b/src/core/record/l1-dedup.test.ts
new file mode 100644
index 00000000..f353ca2a
--- /dev/null
+++ b/src/core/record/l1-dedup.test.ts
@@ -0,0 +1,68 @@
+import { describe, expect, it } from "vitest";
+
+import { batchDedup } from "./l1-dedup.js";
+import type { ExtractedMemory } from "./l1-writer.js";
+import type { IMemoryStore, L1FtsResult } from "../store/types.js";
+import type { LLMRunner } from "../types.js";
+
+describe("batchDedup", () => {
+  it("caps merged priority at the strongest source memory priority", async () => {
+    const newMemory: ExtractedMemory & { record_id: string } = {
+      record_id: "new_1",
+      content: "User asked the agent to only reply OK during diagnostic testing.",
+      type: "instruction",
+      priority: 75,
+      source_message_ids: ["msg_1"],
+      metadata: {},
+      scene_name: "diagnostic testing",
+    };
+
+    const candidate: L1FtsResult = {
+      record_id: "old_1",
+      content: "User asked the assistant not to sound cold during a style adjustment.",
+      type: "instruction",
+      priority: 80,
+      scene_name: "style adjustment",
+      score: 1,
+      timestamp_str: "2026-05-25T08:28:48.298Z",
+      timestamp_start: "2026-05-25T08:28:48.298Z",
+      timestamp_end: "2026-05-25T08:28:48.298Z",
+      session_key: "session-a",
+      session_id: "thread-a",
+      metadata_json: "{}",
+    };
+
+    const vectorStore = {
+      countL1: () => 1,
+      isFtsAvailable: () => true,
+      searchL1Fts: async () => [candidate],
+    } as unknown as IMemoryStore;
+
+    const llmRunner: LLMRunner = {
+      async run() {
+        return JSON.stringify([
+          {
+            record_id: "new_1",
+            action: "merge",
+            target_ids: ["old_1"],
+            merged_content: "User wants extremely concise assistant replies and OK-style confirmations.",
+            merged_type: "instruction",
+            merged_priority: 95,
+            merged_timestamps: ["2026-05-25T08:28:48.298Z"],
+          },
+        ]);
+      },
+    };
+
+    const decisions = await batchDedup({
+      memories: [newMemory],
+      config: {},
+      vectorStore,
+      llmRunner,
+    });
+
+    expect(decisions).toHaveLength(1);
+    expect(decisions[0].action).toBe("merge");
+    expect(decisions[0].merged_priority).toBe(80);
+  });
+});
diff --git a/src/core/record/l1-dedup.ts b/src/core/record/l1-dedup.ts
index 0fb6bf62..bd64f341 100644
--- a/src/core/record/l1-dedup.ts
+++ b/src/core/record/l1-dedup.ts
@@ -178,7 +178,7 @@ async function runLlmJudgment(
     }
 
     const decisions = parseBatchResult(result, memories, logger);
-    return decisions;
+    return capMergedPriorities(decisions, matches, logger);
   } catch (err) {
     logger?.warn?.(
       `${TAG} Batch conflict detection failed, defaulting all to store: ${err instanceof Error ? err.message : String(err)}`,
@@ -386,6 +386,53 @@ function parseBatchResult(
   }
 }
 
+/**
+ * Deterministic guardrail: merge/update can combine evidence, but must not
+ * amplify priority beyond the strongest source memory involved.
+ */
+function capMergedPriorities(
+  decisions: DedupDecision[],
+  matches: CandidateMatch[],
+  logger?: Logger,
+): DedupDecision[] {
+  const matchByRecordId = new Map(matches.map((m) => [m.newMemory.record_id, m]));
+
+  return decisions.map((decision) => {
+    if (
+      (decision.action !== "merge" && decision.action !== "update") ||
+      typeof decision.merged_priority !== "number"
+    ) {
+      return decision;
+    }
+
+    const match = matchByRecordId.get(decision.record_id);
+    if (!match) return decision;
+
+    const targetIds = new Set(decision.target_ids);
+    const sourcePriorities = [
+      match.newMemory.priority,
+      ...match.candidates
+        .filter((candidate) => targetIds.has(candidate.id))
+        .map((candidate) => candidate.priority),
+    ].filter((priority) => Number.isFinite(priority));
+
+    const maxSourcePriority = Math.max(...sourcePriorities);
+    if (!Number.isFinite(maxSourcePriority) || decision.merged_priority <= maxSourcePriority) {
+      return decision;
+    }
+
+    logger?.debug?.(
+      `${TAG} Capping merged priority for ${decision.record_id}: ` +
+      `${decision.merged_priority} → ${maxSourcePriority}`,
+    );
+
+    return {
+      ...decision,
+      merged_priority: maxSourcePriority,
+    };
+  });
+}
+
 /**
  * Fallback: store all memories when parsing fails.
  */