TencentCloud · YOMXXX · Jun 27, 2026
diff --git a/src/core/record/l1-extractor.test.ts b/src/core/record/l1-extractor.test.ts
@@ -0,0 +1,145 @@
+import fs from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+import { extractL1Memories } from "./l1-extractor.js";
+import type { LLMRunner } from "../types.js";
+
+describe("extractL1Memories rule pre-extraction", () => {
+  let tmpDir: string;
+
+  beforeEach(async () => {
+    tmpDir = await fs.mkdtemp(path.join(os.tmpdir(), "tdai-l1-pre-extract-"));
+  });
+
+  afterEach(async () => {
+    await fs.rm(tmpDir, { recursive: true, force: true });
+  });
+
+  it("stores explicit persona statements without calling the LLM", async () => {
+    const llmRunner: LLMRunner = {
+      run: vi.fn(async () => {
+        throw new Error("LLM should not be called for high-confidence persona statements");
+      }),
+    };
+
+    const result = await extractL1Memories({
+      messages: [
+        {
+          id: "msg-1",
+          role: "user",
+          content: "我是 Python 工程师",
+          timestamp: Date.parse("2026-06-01T10:00:00Z"),
+        },
+      ],
+      sessionKey: "session-pre-extract",
+      baseDir: tmpDir,
+      config: {},
+      options: {
+        enableDedup: false,
+        llmRunner,
+      },
+    });
+
+    expect(llmRunner.run).not.toHaveBeenCalled();
+    expect(result.success).toBe(true);
+    expect(result.extractedCount).toBe(1);
+    expect(result.storedCount).toBe(1);
+    expect(result.records[0]).toMatchObject({
+      content: "用户是 Python 工程师。",
+      type: "persona",
+      priority: 80,
+      source_message_ids: ["msg-1"],
+    });
+  });
+
+  it("stores explicit reply instructions without calling the LLM", async () => {
+    const llmRunner: LLMRunner = {
+      run: vi.fn(async () => {
+        throw new Error("LLM should not be called for high-confidence instructions");
+      }),
+    };
+
+    const result = await extractL1Memories({
+      messages: [
+        {
+          id: "msg-1",
+          role: "user",
+          content: "以后请用中文回复",
+          timestamp: Date.parse("2026-06-01T10:00:00Z"),
+        },
+      ],
+      sessionKey: "session-pre-extract",
+      baseDir: tmpDir,
+      config: {},
+      options: {
+        enableDedup: false,
+        llmRunner,
+      },
+    });
+
+    expect(llmRunner.run).not.toHaveBeenCalled();
+    expect(result.success).toBe(true);
+    expect(result.extractedCount).toBe(1);
+    expect(result.storedCount).toBe(1);
+    expect(result.records[0]).toMatchObject({
+      content: "用户要求 AI 以后用中文回复。",
+      type: "instruction",
+      priority: 80,
+      source_message_ids: ["msg-1"],
+    });
+  });
+
+  it("keeps using the LLM when any user message is not covered by direct rules", async () => {
+    const llmRunner: LLMRunner = {
+      run: vi.fn(async () => JSON.stringify([
+        {
+          scene_name: "用户讨论复杂计划",
+          message_ids: ["msg-1", "msg-2"],
+          memories: [
+            {
+              content: "用户正在评估一项复杂迁移计划。",
+              type: "episodic",
+              priority: 70,
+              source_message_ids: ["msg-2"],
+              metadata: {},
+            },
+          ],
+        },
+      ])),
+    };
+
+    const result = await extractL1Memories({
+      messages: [
+        {
+          id: "msg-1",
+          role: "user",
+          content: "我是 Python 工程师",
+          timestamp: Date.parse("2026-06-01T10:00:00Z"),
+        },
+        {
+          id: "msg-2",
+          role: "user",
+          content: "我们今天需要讨论数据库迁移的灰度计划和失败回滚策略。",
+          timestamp: Date.parse("2026-06-01T10:01:00Z"),
+        },
+      ],
+      sessionKey: "session-pre-extract",
+      baseDir: tmpDir,
+      config: {},
+      options: {
+        enableDedup: false,
+        llmRunner,
+      },
+    });
+
+    expect(llmRunner.run).toHaveBeenCalledTimes(1);
+    expect(result.success).toBe(true);
+    expect(result.records).toHaveLength(1);
+    expect(result.records[0]).toMatchObject({
+      content: "用户正在评估一项复杂迁移计划。",
+      type: "episodic",
+      source_message_ids: ["msg-2"],
+    });
+  });
+});
diff --git a/src/core/record/l1-extractor.ts b/src/core/record/l1-extractor.ts
@@ -17,6 +17,7 @@ import { EXTRACT_MEMORIES_SYSTEM_PROMPT, formatExtractionPrompt } from "../promp
 import { batchDedup } from "./l1-dedup.js";
 import { writeMemory, generateMemoryId } from "./l1-writer.js";
 import type { ExtractedMemory, MemoryRecord, MemoryType, DedupDecision } from "./l1-writer.js";
+import { preExtractHighConfidenceMemories } from "./pre-extractor.js";
 import { CleanContextRunner } from "../../utils/clean-context-runner.js";
 import { sanitizeJsonForParse, shouldExtractL1 } from "../../utils/sanitize.js";
 import type { IMemoryStore } from "../store/types.js";
@@ -148,44 +149,57 @@ export async function extractL1Memories(params: {
 
   logger?.debug?.(`${TAG} Extracting from ${newMessages.length} new messages (+ ${backgroundMessages.length} background) [${qualifiedMessages.length} qualified from ${messages.length} input]`);
 
-  // Step 1: LLM extraction (scene segmentation + memory extraction)
-  let scenes: SceneSegment[];
-  try {
-    scenes = await callLlmExtraction({
-      newMessages,
-      backgroundMessages,
-      previousSceneName: options.previousSceneName,
-      config,
-      logger,
-      model: options.model,
-      llmRunner: options.llmRunner,
-    });
-    logger?.debug?.(`${TAG} LLM detected ${scenes.length} scene(s)`);
-  } catch (err) {
-    logger?.error(`${TAG} LLM extraction failed: ${err instanceof Error ? err.message : String(err)}`);
-    return { success: false, extractedCount: 0, storedCount: 0, records: [], sceneNames: [] };
-  }
-
-  // Flatten all memories across scenes
   const allExtracted: ExtractedMemory[] = [];
   const sceneNames: string[] = [];
 
-  for (const scene of scenes) {
-    sceneNames.push(scene.scene_name);
-    for (const mem of scene.memories) {
-      const memType = normalizeType(mem.type);
-      if (!memType) {
-        logger?.warn?.(`${TAG} Skipping memory with invalid type "${mem.type}"`);
-        continue;
+  const preExtraction = preExtractHighConfidenceMemories(newMessages);
+  if (preExtraction.canBypassLlm) {
+    for (const memory of preExtraction.memories) {
+      allExtracted.push(memory);
+      if (!sceneNames.includes(memory.scene_name)) {
+        sceneNames.push(memory.scene_name);
       }
-      allExtracted.push({
-        content: mem.content,
-        type: memType,
-        priority: typeof mem.priority === "number" ? mem.priority : 50,
-        source_message_ids: Array.isArray(mem.source_message_ids) ? mem.source_message_ids : [],
-        metadata: mem.metadata ?? {},
-        scene_name: scene.scene_name,
+    }
+    logger?.debug?.(
+      `${TAG} Rule pre-extraction bypassed LLM: ${allExtracted.length} high-confidence memor${allExtracted.length === 1 ? "y" : "ies"}`,
+    );
+  } else {
+    // Step 1: LLM extraction (scene segmentation + memory extraction)
+    let scenes: SceneSegment[];
+    try {
+      scenes = await callLlmExtraction({
+        newMessages,
+        backgroundMessages,
+        previousSceneName: options.previousSceneName,
+        config,
+        logger,
+        model: options.model,
+        llmRunner: options.llmRunner,
       });
+      logger?.debug?.(`${TAG} LLM detected ${scenes.length} scene(s)`);
+    } catch (err) {
+      logger?.error(`${TAG} LLM extraction failed: ${err instanceof Error ? err.message : String(err)}`);
+      return { success: false, extractedCount: 0, storedCount: 0, records: [], sceneNames: [] };
+    }
+
+    // Flatten all memories across scenes
+    for (const scene of scenes) {
+      sceneNames.push(scene.scene_name);
+      for (const mem of scene.memories) {
+        const memType = normalizeType(mem.type);
+        if (!memType) {
+          logger?.warn?.(`${TAG} Skipping memory with invalid type "${mem.type}"`);
+          continue;
+        }
+        allExtracted.push({
+          content: mem.content,
+          type: memType,
+          priority: typeof mem.priority === "number" ? mem.priority : 50,
+          source_message_ids: Array.isArray(mem.source_message_ids) ? mem.source_message_ids : [],
+          metadata: mem.metadata ?? {},
+          scene_name: scene.scene_name,
+        });
+      }
     }
   }
 

diff --git a/src/core/record/pre-extractor.ts b/src/core/record/pre-extractor.ts
@@ -0,0 +1,122 @@
+import type { ConversationMessage } from "../conversation/l0-recorder.js";
+import type { ExtractedMemory } from "./l1-writer.js";
+
+export interface PreExtractionResult {
+  memories: ExtractedMemory[];
+  canBypassLlm: boolean;
+}
+
+type RuleMatch = Omit<ExtractedMemory, "source_message_ids">;
+
+const MAX_DIRECT_MESSAGE_CHARS = 80;
+
+export function preExtractHighConfidenceMemories(messages: ConversationMessage[]): PreExtractionResult {
+  const userMessages = messages.filter((message) => message.role === "user");
+  const memories: ExtractedMemory[] = [];
+  const coveredUserIds = new Set<string>();
+
+  for (const message of userMessages) {
+    const match = matchHighConfidenceRule(message.content);
+    if (!match) continue;
+    memories.push({
+      ...match,
+      source_message_ids: [message.id],
+    });
+    coveredUserIds.add(message.id);
+  }
+
+  return {
+    memories,
+    canBypassLlm: memories.length > 0 && userMessages.every((message) => coveredUserIds.has(message.id)),
+  };
+}
+
+function matchHighConfidenceRule(raw: string): RuleMatch | undefined {
+  const text = normalizeDirectText(raw);
+  if (!text || text.length > MAX_DIRECT_MESSAGE_CHARS || /[\n\r]/.test(raw)) {
+    return undefined;
+  }
+
+  return matchPersonaRule(text) ?? matchInstructionRule(text);
+}
+
+function matchPersonaRule(text: string): RuleMatch | undefined {
+  const identity = text.match(/^我(?:是一名|是一个|是一位|是)\s*(.{2,40})$/u);
+  if (identity) {
+    return {
+      content: `用户是 ${identity[1].trim()}。`,
+      type: "persona",
+      priority: 80,
+      scene_name: "用户介绍个人身份信息",
+      metadata: {},
+    };
+  }
+
+  const occupation = text.match(/^我的(?:职业|工作|岗位)(?:是|为)\s*(.{2,40})$/u);
+  if (occupation) {
+    return {
+      content: `用户的职业是 ${occupation[1].trim()}。`,
+      type: "persona",
+      priority: 80,
+      scene_name: "用户介绍个人身份信息",
+      metadata: {},
+    };
+  }
+
+  const preference = text.match(/^我(喜欢|偏好|擅长)\s*(.{2,40})$/u);
+  if (preference) {
+    return {
+      content: `用户${preference[1]} ${preference[2].trim()}。`,
+      type: "persona",
+      priority: 70,
+      scene_name: "用户介绍个人偏好",
+      metadata: {},
+    };
+  }
+
+  return undefined;
+}
+
+function matchInstructionRule(text: string): RuleMatch | undefined {
+  const language = text.match(/^以后(?:请|都|要)?\s*(?:用|使用)\s*(.{1,20}?)(?:回复|回答)$/u);
+  if (language) {
+    return {
+      content: `用户要求 AI 以后用${language[1].trim()}回复。`,
+      type: "instruction",
+      priority: 80,
+      scene_name: "用户设置 AI 回复偏好",
+      metadata: {},
+    };
+  }
+
+  const futureDirective = text.match(/^以后(?:请|都|要)?\s*(不要|别|保持|尽量|必须|直接)\s*(.{2,40})$/u);
+  if (futureDirective) {
+    return {
+      content: `用户要求 AI 以后${futureDirective[1]}${futureDirective[2].trim()}。`,
+      type: "instruction",
+      priority: 80,
+      scene_name: "用户设置 AI 回复偏好",
+      metadata: {},
+    };
+  }
+
+  const fromNow = text.match(/^从现在开始(?:请|都|要)?\s*(.{2,50})$/u);
+  if (fromNow) {
+    return {
+      content: `用户要求 AI 从现在开始${fromNow[1].trim()}。`,
+      type: "instruction",
+      priority: 80,
+      scene_name: "用户设置 AI 回复偏好",
+      metadata: {},
+    };
+  }
+
+  return undefined;
+}
+
+function normalizeDirectText(text: string): string {
+  return text
+    .trim()
+    .replace(/[。.!！]+$/u, "")
+    .replace(/\s+/g, " ");
+}