memory-lancedb-pro 1.0.26 → 1.1.0-beta.2

package/CHANGELOG-v1.1.0.md

@@ -0,0 +1,227 @@
+# memory-lancedb-pro v1.1.0 — 智能记忆增强
+
+> **日期**: 2026-03-03  
+> **作者**: CJY  
+> **概述**: 基于对 AI Agent 记忆系统的深入理解，对记忆的写入质量、生命周期管理和去重能力进行了全面改进与完善
+
+---
+
+## 一、改进动机
+
+原有记忆系统在**检索侧**表现优异（Vector+BM25 混合检索、cross-encoder 重排序、多维评分），但在以下方面存在提升空间：
+
+- **记忆写入质量**：依赖正则表达式触发捕获，容易漏捕有价值信息或误捕噪声
+- **记忆结构层次**：扁平文本存储，缺乏分层索引能力
+- **记忆生命周期**：简单时间衰减，无法模拟人类记忆的遗忘与强化规律
+- **去重能力**：仅基于向量相似度的粗粒度去重，缺乏语义级判断
+
+本次改进针对这三个维度进行了系统性增强。
+
+---
+
+## 二、变更摘要
+
+| 改进维度     | 核心变更                                  | 效果                               |
+| ------------ | ----------------------------------------- | ---------------------------------- |
+| 智能提取     | LLM 驱动的 6 类别提取 + L0/L1/L2 分层存储 | 记忆写入更精准、结构更丰富         |
+| 生命周期管理 | Weibull 衰减模型 + 三层晋升/降级          | 重要记忆持久保留，过时记忆自然淡化 |
+| 智能去重     | 向量预过滤 + LLM 语义决策                 | 避免冗余记忆，支持信息演化合并     |
+
+---
+
+## 三、新增文件
+
+### 1. `src/memory-categories.ts` — 6 类别分类系统
+
+设计了语义明确的记忆分类体系，将记忆分为两大类六小类：
+
+- **用户记忆**：`profile`（身份属性）、`preferences`（偏好习惯）、`entities`（持续存在的实体）、`events`（发生的事件）
+- **Agent 记忆**：`cases`（问题-解决方案对）、`patterns`（可复用的处理流程）
+
+每个类别有不同的合并策略：
+
+- `profile` → 始终合并（用户身份信息持续累积）
+- `preferences` / `entities` / `patterns` → 支持智能合并
+- `events` / `cases` → 仅新增或跳过（独立记录，保留历史完整性）
+
+---
+
+### 2. `src/llm-client.ts` — LLM 客户端
+
+封装了 LLM 调用接口，专注于结构化 JSON 输出：
+
+- 复用现有 OpenAI SDK 依赖，零新增包
+- 内置 JSON 容错解析：支持 markdown 代码块包裹和平衡大括号提取
+- 低温度 (0.1) 保证输出一致性
+- 30 秒超时保护，失败时优雅降级
+
+---
+
+### 3. `src/extraction-prompts.ts` — 记忆提取提示模板
+
+精心设计了 3 个提示模板：
+
+| 函数                      | 用途                                                |
+| ------------------------- | --------------------------------------------------- |
+| `buildExtractionPrompt()` | 从对话中提取 6 类别 L0/L1/L2 记忆，含 few-shot 示例 |
+| `buildDedupPrompt()`      | CREATE / MERGE / SKIP 去重决策                      |
+| `buildMergePrompt()`      | 将新旧记忆合并为三层结构                            |
+
+提取提示包含完整的记忆价值判断标准、类别决策逻辑表、常见混淆澄清规则和 6 个 few-shot 示例。
+
+---
+
+### 4. `src/smart-extractor.ts` — 智能提取管线
+
+实现了完整的 LLM 驱动提取流水线：
+
+```
+对话文本 → LLM 提取 → 候选记忆 → 向量去重 → LLM 决策 → 持久化
+```
+
+核心设计：
+
+- **两阶段去重**：先用向量相似度（阈值 0.7）快速筛选候选，再用 LLM 进行语义级判断
+- **类别感知合并**：不同类别应用不同合并策略
+- **L0/L1/L2 三层存储**：L0 一句话索引用于检索注入，L1 结构化摘要用于精读，L2 完整叙述用于深度回顾
+- **向后兼容**：新增的 6 类别自动映射到已有的 5 类别存储，L0/L1/L2 存储在 metadata JSON 中
+- **按类别设定重要度**：profile (0.9) > patterns (0.85) > cases/preferences (0.8) > entities (0.7) > events (0.6)
+
+---
+
+### 5. `src/decay-engine.ts` — Weibull 衰减引擎
+
+基于认知心理学中的记忆遗忘曲线研究，实现了复合衰减模型：
+
+**复合分数 = 时效权重 × 时效 + 频率权重 × 频率 + 内在权重 × 内在价值**
+
+三个分量：
+
+| 分量                     | 机制                              | 含义                   |
+| ------------------------ | --------------------------------- | ---------------------- |
+| **时效 (recency)**       | Weibull 拉伸指数衰减 `exp(-λt^β)` | 越久远的记忆衰减越快   |
+| **频率 (frequency)**     | 对数饱和曲线 + 时间加权           | 越常被访问的记忆越活跃 |
+| **内在价值 (intrinsic)** | `importance × confidence`         | 高价值记忆天然抵抗遗忘 |
+
+层级特定的衰减形状 (β 参数)：
+
+- **Core** (β=0.8)：亚指数衰减 → 遗忘极慢，衰减地板 0.9
+- **Working** (β=1.0)：标准指数衰减，衰减地板 0.7
+- **Peripheral** (β=1.3)：超指数衰减 → 遗忘加速，衰减地板 0.5
+
+关键特性：
+
+- **重要性调制半衰期**：`effectiveHL = halfLife × exp(μ × importance)`，重要记忆持续更久
+- **搜索结果加权**：检索时自动应用衰减加权，让活跃记忆排名更高
+- **过期识别**：识别 composite < 0.3 的过期记忆
+
+---
+
+### 6. `src/tier-manager.ts` — 三层晋升/降级管理器
+
+模拟人类记忆的多级存储模型：
+
+```
+Peripheral（外围） ⟷ Working（工作） ⟷ Core（核心）
+```
+
+**晋升条件**：
+
+| 方向                 | 条件                                            |
+| -------------------- | ----------------------------------------------- |
+| Peripheral → Working | 访问次数 ≥ 3 且 衰减分数 ≥ 0.4                  |
+| Working → Core       | 访问次数 ≥ 10 且 衰减分数 ≥ 0.7 且 重要度 ≥ 0.8 |
+
+**降级条件**：
+
+| 方向                 | 条件                                             |
+| -------------------- | ------------------------------------------------ |
+| Working → Peripheral | 衰减分数 < 0.15 或（年龄 > 60 天且访问次数 < 3） |
+| Core → Working       | 衰减分数 < 0.15 且 访问次数 < 3（极少触发）      |
+
+---
+
+## 四、修改文件
+
+### `index.ts` — 插件入口
+
+#### 新增配置项
+
+```typescript
+smartExtraction?: boolean;    // 是否启用 LLM 智能提取（默认 true）
+llm?: {
+  apiKey?: string;            // LLM API Key（默认复用 embedding.apiKey）
+  model?: string;             // LLM 模型（默认 gpt-4o-mini）
+  baseURL?: string;           // LLM API 端点
+};
+extractMinMessages?: number;  // 最少消息数才触发提取（默认 4）
+extractMaxChars?: number;     // 送入 LLM 的最大字符数（默认 8000）
+```
+
+#### `agent_end` 钩子改进
+
+- 当 `smartExtraction` 启用时，优先使用 SmartExtractor 进行 LLM 6 类别提取
+- 当消息数不足或 SmartExtractor 未初始化时，降级回原有正则触发逻辑
+- 提取完成后输出统计日志：`smart-extracted N created, M merged, K skipped`
+
+#### `before_agent_start` 钩子改进
+
+- 注入的记忆上下文现在显示 L0 摘要而非原始文本
+- 新增 6 类别标签（如 `[preferences:global]`）
+- 新增层级标记（`[C]`ore / `[W]`orking / `[P]`eripheral）
+
+---
+
+## 五、配置指南
+
+### 最简配置（复用已有 API Key）
+
+```json
+{
+  "embedding": {
+    "apiKey": "${OPENAI_API_KEY}",
+    "model": "text-embedding-3-small"
+  },
+  "smartExtraction": true
+}
+```
+
+### 完整配置
+
+```json
+{
+  "embedding": {
+    "apiKey": "${OPENAI_API_KEY}",
+    "model": "text-embedding-3-small"
+  },
+  "smartExtraction": true,
+  "llm": {
+    "apiKey": "${OPENAI_API_KEY}",
+    "model": "gpt-4o-mini",
+    "baseURL": "https://api.openai.com/v1"
+  },
+  "extractMinMessages": 4,
+  "extractMaxChars": 8000
+}
+```
+
+### 禁用智能提取
+
+```json
+{
+  "smartExtraction": false
+}
+```
+
+---
+
+## 六、向后兼容性
+
+| 方面           | 兼容方式                                       |
+| -------------- | ---------------------------------------------- |
+| LanceDB Schema | 新字段存储在 `metadata` JSON 中，不修改表结构  |
+| 记忆类别       | 新 6 类别自动映射到原有 5 类别                 |
+| 混合检索       | Vector+BM25 检索管线完全保留                   |
+| 去重逻辑       | 仅在 `smartExtraction: true` 时生效            |
+| 已有数据       | 旧记忆正常读取，新记忆额外携带 L0/L1/L2 元数据 |
+| 配置           | 全部新增配置项均有默认值，零配置即可使用       |

package/CHANGELOG.md

@@ -1,5 +1,28 @@
 # Changelog
 
+## 1.1.0-beta.2 (Smart Memory Beta + Access Reinforcement)
+
+This is a **beta** release published under the npm dist-tag **`beta`** (it does not affect the stable `latest` channel).
+
+Highlights:
+- **Smart Extraction (LLM-powered)**: 6-category extraction with L0/L1/L2 metadata (falls back to regex capture when disabled or init fails)
+- **Lifecycle scoring integrated into retrieval**: decay-based score adjustment + tier floors
+- **Tier transitions (best-effort)**: bounded metadata write-backs for top results (tier / access stats)
+- **Access reinforcement for time decay**: frequently *manually recalled* memories decay more slowly (spaced-repetition style)
+  - Adds `AccessTracker` with debounced metadata write-back (accessCount / lastAccessedAt)
+  - Adds retrieval config: `reinforcementFactor` (default: 0.5) and `maxHalfLifeMultiplier` (default: 3)
+
+Notes:
+- Access reinforcement is gated to manual recall (`source: \"manual\"`) to avoid auto-recall strengthening noise.
+
+---
+
+## 1.1.0-beta.1 (Smart Memory Beta)
+
+- Initial beta with Smart Extraction + lifecycle components (decay engine + tier manager)
+
+---
+
 ## 1.0.26
 
 **Access Reinforcement for Time Decay**

package/README.md

@@ -52,6 +52,82 @@ The built-in `memory-lancedb` plugin in OpenClaw provides basic vector search. *
 
 ---
 
+## 🧪 Beta: Smart Memory v1.1.0
+
+> **Status**: Beta — available on npm under the `beta` dist-tag. Stable users on `latest` are not affected.
+
+The `dev/smart-memory-v1.1.0` branch introduces three major enhancements to the memory write & retrieval pipeline:
+
+### What's New
+
+| Feature | Description |
+|---------|-------------|
+| **Smart Extraction** | LLM-powered 6-category extraction (profile, preferences, entities, events, cases, patterns) with L0/L1/L2 layered metadata. Falls back to regex capture when disabled or LLM init fails. |
+| **Lifecycle Scoring** | Weibull decay model integrated into retrieval — scores are adjusted by `max(tierFloor, decayComposite)` so frequently-accessed and high-importance memories rank higher. |
+| **Tier Management** | Three-tier system (Core → Working → Peripheral) with automatic promotion/demotion based on access frequency, composite score, and importance. |
+
+### Install the Beta
+
+```bash
+npm i memory-lancedb-pro@beta
+```
+
+Or pin the exact version:
+
+```bash
+npm i memory-lancedb-pro@1.1.0-beta.1
+```
+
+### Configuration
+
+Smart extraction is **enabled by default**. It reuses your existing embedding API key for LLM calls (or you can configure a separate LLM endpoint):
+
+```json
+{
+  "plugins.entries.memory-lancedb-pro": {
+    "config": {
+      "smartExtraction": true,
+      "llm": {
+        "apiKey": "${OPENAI_API_KEY}",
+        "model": "gpt-4o-mini",
+        "baseURL": "https://api.openai.com/v1"
+      },
+      "extractMinMessages": 4,
+      "extractMaxChars": 8000
+    }
+  }
+}
+```
+
+| Config Key | Default | Description |
+|------------|---------|-------------|
+| `smartExtraction` | `true` | Enable/disable LLM-powered extraction |
+| `llm.apiKey` | *(embedding apiKey)* | API key for extraction LLM |
+| `llm.model` | `gpt-4o-mini` | LLM model for extraction & dedup |
+| `llm.baseURL` | *(embedding baseURL)* | Base URL for LLM API |
+| `extractMinMessages` | `4` | Min conversation messages before extraction triggers |
+| `extractMaxChars` | `8000` | Max conversation chars to process |
+
+### New Files
+
+| File | Purpose |
+|------|---------|
+| `src/smart-extractor.ts` | LLM extraction pipeline: conversation → extract → dedup → persist |
+| `src/extraction-prompts.ts` | Prompt templates for extraction, dedup, and merge |
+| `src/llm-client.ts` | OpenAI-compatible LLM client with JSON parsing |
+| `src/memory-categories.ts` | 6-category classification system + merge strategies |
+| `src/decay-engine.ts` | Weibull stretched-exponential decay with tier-specific beta |
+| `src/tier-manager.ts` | Three-tier promotion/demotion lifecycle manager |
+
+### Feedback
+
+This is a beta release — please report issues or share feedback at:
+- [GitHub Issues](https://github.com/win4r/memory-lancedb-pro/issues)
+
+To revert to stable: `npm i memory-lancedb-pro@latest`
+
+---
+
 ## Architecture
 
 ```
@@ -92,6 +168,12 @@ The built-in `memory-lancedb` plugin in OpenClaw provides basic vector search. *
 | `src/noise-filter.ts` | Noise filter. Filters out agent refusals, meta-questions, greetings, and low-quality content |
 | `src/adaptive-retrieval.ts` | Adaptive retrieval. Determines whether a query needs memory retrieval (skips greetings, slash commands, simple confirmations, emoji) |
 | `src/migrate.ts` | Migration tool. Migrates data from the built-in `memory-lancedb` plugin to Pro |
+| `src/smart-extractor.ts` | *(Beta)* LLM-powered 6-category extraction pipeline with L0/L1/L2 layered storage |
+| `src/extraction-prompts.ts` | *(Beta)* Prompt templates for memory extraction, dedup decisions, and merge |
+| `src/llm-client.ts` | *(Beta)* OpenAI-compatible LLM client wrapper with JSON fence parsing |
+| `src/memory-categories.ts` | *(Beta)* 6-category classification (profile, preferences, entities, events, cases, patterns) |
+| `src/decay-engine.ts` | *(Beta)* Weibull decay model with importance-modulated half-life and tier-specific beta |
+| `src/tier-manager.ts` | *(Beta)* Three-tier (Core/Working/Peripheral) promotion/demotion lifecycle manager |
 
 ---
 

package/README_CN.md

@@ -52,6 +52,82 @@ OpenClaw 内置的 `memory-lancedb` 插件仅提供基本的向量搜索。**mem
 
 ---
 
+## 🧪 Beta：智能记忆 v1.1.0
+
+> **状态**：Beta 版 — 通过 npm `beta` dist-tag 发布，不影响 `latest` 稳定通道。
+
+`dev/smart-memory-v1.1.0` 分支为记忆写入和检索管线引入了三大增强：
+
+### 新功能
+
+| 功能 | 说明 |
+|------|------|
+| **智能提取** | LLM 驱动的 6 类别提取（profile、preferences、entities、events、cases、patterns），支持 L0/L1/L2 分层元数据。禁用或 LLM 初始化失败时回退到正则捕获。 |
+| **生命周期评分** | Weibull 衰减模型集成到检索中 — 分数通过 `max(tierFloor, decayComposite)` 调整，高频访问和高重要性的记忆排名更靠前。 |
+| **分层管理** | 三层系统（Core → Working → Peripheral），根据访问频率、复合得分和重要性自动晋升/降级。 |
+
+### 安装 Beta 版
+
+```bash
+npm i memory-lancedb-pro@beta
+```
+
+或指定精确版本：
+
+```bash
+npm i memory-lancedb-pro@1.1.0-beta.1
+```
+
+### 配置
+
+智能提取**默认开启**。它复用你现有的 embedding API key 进行 LLM 调用（也可以单独配置 LLM 端点）：
+
+```json
+{
+  "plugins.entries.memory-lancedb-pro": {
+    "config": {
+      "smartExtraction": true,
+      "llm": {
+        "apiKey": "${OPENAI_API_KEY}",
+        "model": "gpt-4o-mini",
+        "baseURL": "https://api.openai.com/v1"
+      },
+      "extractMinMessages": 4,
+      "extractMaxChars": 8000
+    }
+  }
+}
+```
+
+| 配置项 | 默认值 | 说明 |
+|--------|--------|------|
+| `smartExtraction` | `true` | 启用/禁用 LLM 驱动提取 |
+| `llm.apiKey` | *(embedding apiKey)* | 提取 LLM 的 API key |
+| `llm.model` | `gpt-4o-mini` | 提取和去重使用的 LLM 模型 |
+| `llm.baseURL` | *(embedding baseURL)* | LLM API 的 Base URL |
+| `extractMinMessages` | `4` | 触发提取的最少对话消息数 |
+| `extractMaxChars` | `8000` | 处理提取的最大对话字符数 |
+
+### 新增文件
+
+| 文件 | 用途 |
+|------|------|
+| `src/smart-extractor.ts` | LLM 提取管线：对话 → 提取 → 去重 → 持久化 |
+| `src/extraction-prompts.ts` | 提取、去重和合并的提示词模板 |
+| `src/llm-client.ts` | OpenAI 兼容 LLM 客户端，含 JSON 解析 |
+| `src/memory-categories.ts` | 6 类别分类系统 + 合并策略 |
+| `src/decay-engine.ts` | Weibull 拉伸指数衰减模型 |
+| `src/tier-manager.ts` | 三层晋升/降级生命周期管理器 |
+
+### 反馈
+
+这是 beta 版本 — 欢迎在以下地方报告问题或分享反馈：
+- [GitHub Issues](https://github.com/win4r/memory-lancedb-pro/issues)
+
+回退到稳定版：`npm i memory-lancedb-pro@latest`
+
+---
+
 ## 架构概览
 
 ```
@@ -92,6 +168,12 @@ OpenClaw 内置的 `memory-lancedb` 插件仅提供基本的向量搜索。**mem
 | `src/noise-filter.ts` | 噪声过滤器。过滤 Agent 拒绝回复、Meta 问题、寒暄等低质量记忆 |
 | `src/adaptive-retrieval.ts` | 自适应检索。判断 query 是否需要触发记忆检索（跳过问候、命令、简单确认等） |
 | `src/migrate.ts` | 迁移工具。从旧版 `memory-lancedb` 插件迁移数据到 Pro 版 |
+| `src/smart-extractor.ts` | *(Beta)* LLM 驱动的 6 类别提取管线，L0/L1/L2 分层存储 |
+| `src/extraction-prompts.ts` | *(Beta)* 记忆提取、去重决策和合并的提示词模板 |
+| `src/llm-client.ts` | *(Beta)* OpenAI 兼容 LLM 客户端封装，含 JSON 围栏解析 |
+| `src/memory-categories.ts` | *(Beta)* 6 类别分类（profile、preferences、entities、events、cases、patterns） |
+| `src/decay-engine.ts` | *(Beta)* Weibull 衰减模型，重要性调制半衰期 + 分层 beta |
+| `src/tier-manager.ts` | *(Beta)* 三层（Core/Working/Peripheral）晋升/降级生命周期管理器 |
 
 ---
 

package/index.ts

@@ -20,6 +20,12 @@ import { shouldSkipRetrieval } from "./src/adaptive-retrieval.js";
 import { AccessTracker } from "./src/access-tracker.js";
 import { createMemoryCLI } from "./cli.js";
 
+// Import smart extraction & lifecycle components
+import { SmartExtractor } from "./src/smart-extractor.js";
+import { createLlmClient } from "./src/llm-client.js";
+import { createDecayEngine, DEFAULT_DECAY_CONFIG } from "./src/decay-engine.js";
+import { createTierManager, DEFAULT_TIER_CONFIG } from "./src/tier-manager.js";
+
 // ============================================================================
 // Configuration & Types
 // ============================================================================
@@ -27,7 +33,7 @@ import { createMemoryCLI } from "./cli.js";
 interface PluginConfig {
   embedding: {
     provider: "openai-compatible";
-    apiKey: string;
+    apiKey: string | string[];
     model?: string;
     baseURL?: string;
     dimensions?: number;
@@ -60,6 +66,15 @@ interface PluginConfig {
     reinforcementFactor?: number;
     maxHalfLifeMultiplier?: number;
   };
+  // Smart extraction config (Phase 1: from epro-memory)
+  smartExtraction?: boolean;
+  llm?: {
+    apiKey?: string;
+    model?: string;
+    baseURL?: string;
+  };
+  extractMinMessages?: number;
+  extractMaxChars?: number;
   scopes?: {
     default?: string;
     definitions?: Record<string, { description: string }>;
@@ -398,10 +413,19 @@ const memoryLanceDBProPlugin = {
       taskPassage: config.embedding.taskPassage,
       normalized: config.embedding.normalized,
     });
-    const retriever = createRetriever(store, embedder, {
-      ...DEFAULT_RETRIEVAL_CONFIG,
-      ...config.retrieval,
-    });
+    // Initialize decay engine + tier manager (lifecycle scoring)
+    const decayEngine = createDecayEngine(DEFAULT_DECAY_CONFIG);
+    const tierManager = createTierManager(DEFAULT_TIER_CONFIG);
+
+    const retriever = createRetriever(
+      store,
+      embedder,
+      {
+        ...DEFAULT_RETRIEVAL_CONFIG,
+        ...config.retrieval,
+      },
+      { decayEngine, tierManager },
+    );
 
     // Access reinforcement tracker (debounced write-back)
     const accessTracker = new AccessTracker({
@@ -414,10 +438,46 @@ const memoryLanceDBProPlugin = {
     const scopeManager = createScopeManager(config.scopes);
     const migrator = createMigrator(store);
 
+    // Initialize smart extraction (Phase 1: from epro-memory)
+    let smartExtractor: SmartExtractor | null = null;
+    if (config.smartExtraction !== false) {
+      try {
+        const embeddingKey = Array.isArray(config.embedding.apiKey)
+          ? config.embedding.apiKey[0]
+          : config.embedding.apiKey;
+        const llmApiKey = config.llm?.apiKey
+          ? resolveEnvVars(config.llm.apiKey)
+          : resolveEnvVars(embeddingKey);
+        const llmBaseURL = config.llm?.baseURL
+          ? resolveEnvVars(config.llm.baseURL)
+          : config.embedding.baseURL;
+        const llmModel = config.llm?.model || "gpt-4o-mini";
+
+        const llmClient = createLlmClient({
+          apiKey: llmApiKey,
+          model: llmModel,
+          baseURL: llmBaseURL,
+          timeoutMs: 30000,
+        });
+
+        smartExtractor = new SmartExtractor(store, embedder, llmClient, {
+          user: "User",
+          extractMinMessages: config.extractMinMessages ?? 4,
+          extractMaxChars: config.extractMaxChars ?? 8000,
+          defaultScope: config.scopes?.default ?? "global",
+          log: (msg: string) => api.logger.info(msg),
+        });
+
+        api.logger.info("memory-lancedb-pro: smart extraction enabled (LLM model: " + llmModel + ")");
+      } catch (err) {
+        api.logger.warn(`memory-lancedb-pro: smart extraction init failed, falling back to regex: ${String(err)}`);
+      }
+    }
+
     const pluginVersion = getPluginVersion();
 
     api.logger.info(
-      `memory-lancedb-pro@${pluginVersion}: plugin registered (db: ${resolvedDbPath}, model: ${config.embedding.model || "text-embedding-3-small"})`,
+      `memory-lancedb-pro@${pluginVersion}: plugin registered (db: ${resolvedDbPath}, model: ${config.embedding.model || "text-embedding-3-small"}, smartExtraction: ${smartExtractor ? "ON" : "OFF"})`,
     );
 
     // ========================================================================
@@ -484,11 +544,19 @@ const memoryLanceDBProPlugin = {
             return;
           }
 
+          // Format with L0 abstracts grouped by category when available
           const memoryContext = results
-            .map(
-              (r) =>
-                `- [${r.entry.category}:${r.entry.scope}] ${sanitizeForContext(r.entry.text)} (${(r.score * 100).toFixed(0)}%${r.sources?.bm25 ? ", vector+BM25" : ""}${r.sources?.reranked ? "+reranked" : ""})`,
-            )
+            .map((r) => {
+              let metaObj: Record<string, unknown> = {};
+              try {
+                metaObj = JSON.parse(r.entry.metadata || "{}");
+              } catch {}
+              const displayCategory = (metaObj.memory_category as string) || r.entry.category;
+              const displayTier = (metaObj.tier as string) || "";
+              const tierPrefix = displayTier ? `[${displayTier.charAt(0).toUpperCase()}]` : "";
+              const abstract = (metaObj.l0_abstract as string) || r.entry.text;
+              return `- ${tierPrefix}[${displayCategory}:${r.entry.scope}] ${sanitizeForContext(abstract)} (${(r.score * 100).toFixed(0)}%${r.sources?.bm25 ? ", vector+BM25" : ""}${r.sources?.reranked ? "+reranked" : ""})`;
+            })
             .join("\n");
 
           api.logger.info?.(
@@ -561,7 +629,29 @@ const memoryLanceDBProPlugin = {
             }
           }
 
-          // Filter for capturable content
+          // ----------------------------------------------------------------
+          // Smart Extraction (Phase 1: LLM-powered 6-category extraction)
+          // ----------------------------------------------------------------
+          if (smartExtractor) {
+            const minMessages = config.extractMinMessages ?? 4;
+            if (texts.length >= minMessages) {
+              const conversationText = texts.join("\n");
+              const sessionKey = (event as any).sessionKey || "unknown";
+              const stats = await smartExtractor.extractAndPersist(
+                conversationText, sessionKey,
+              );
+              if (stats.created > 0 || stats.merged > 0) {
+                api.logger.info(
+                  `memory-lancedb-pro: smart-extracted ${stats.created} created, ${stats.merged} merged, ${stats.skipped} skipped for agent ${agentId}`
+                );
+              }
+              return; // Smart extraction handled everything
+            }
+          }
+
+          // ----------------------------------------------------------------
+          // Fallback: regex-triggered capture (original logic)
+          // ----------------------------------------------------------------
           const toCapture = texts.filter((text) => text && shouldCapture(text));
           if (toCapture.length === 0) {
             return;
@@ -934,6 +1024,11 @@ function parsePluginConfig(value: unknown): PluginConfig {
       typeof cfg.retrieval === "object" && cfg.retrieval !== null
         ? (cfg.retrieval as any)
         : undefined,
+    // Smart extraction config (Phase 1)
+    smartExtraction: cfg.smartExtraction !== false, // Default ON
+    llm: typeof cfg.llm === "object" && cfg.llm !== null ? (cfg.llm as any) : undefined,
+    extractMinMessages: parsePositiveInt(cfg.extractMinMessages) ?? 4,
+    extractMaxChars: parsePositiveInt(cfg.extractMaxChars) ?? 8000,
     scopes:
       typeof cfg.scopes === "object" && cfg.scopes !== null
         ? (cfg.scopes as any)