feat: implement ZCLAW Agent Intelligence Evolution Phase 1-3
Phase 1: Persistent Memory + Identity Dynamic Evolution - agent-memory.ts: MemoryManager with localStorage persistence, keyword search, deduplication, importance scoring, pruning, markdown export - agent-identity.ts: AgentIdentityManager with per-agent SOUL/AGENTS/USER.md, change proposals with approval workflow, snapshot rollback - memory-extractor.ts: Rule-based conversation memory extraction (Phase 1), LLM extraction prompt ready for Phase 2 - MemoryPanel.tsx: Memory browsing UI with search, type filter, delete, export (integrated as 4th tab in RightPanel) Phase 2: Context Governance - context-compactor.ts: Token estimation, threshold monitoring (soft/hard), memory flush before compaction, rule-based summarization - chatStore integration: auto-compact when approaching token limits Phase 3: Proactive Intelligence + Self-Reflection - heartbeat-engine.ts: Periodic checks (pending tasks, memory health, idle greeting), quiet hours, proactivity levels (silent/light/standard/autonomous) - reflection-engine.ts: Pattern analysis from memory corpus, improvement suggestions, identity change proposals, meta-memory creation Chat Flow Integration (chatStore.ts): - Pre-send: context compaction check -> memory search -> identity system prompt injection - Post-complete: async memory extraction -> reflection conversation tracking -> auto-trigger reflection Tests: 274 passing across 12 test files - agent-memory.test.ts: 42 tests - context-compactor.test.ts: 23 tests - heartbeat-reflection.test.ts: 28 tests - chatStore.test.ts: 11 tests (no regressions) Refs: ZCLAW_AGENT_INTELLIGENCE_EVOLUTION.md updated with implementation progress
This commit is contained in:
iven
319
desktop/src/lib/memory-extractor.ts
Normal file
319
desktop/src/lib/memory-extractor.ts
Normal file
@@ -0,0 +1,319 @@
|
||||
/**
|
||||
* Memory Extractor - Automatically extract memorable information from conversations
|
||||
*
|
||||
* Uses LLM to analyze completed conversations and extract:
|
||||
* - Facts the user shared
|
||||
* - User preferences discovered
|
||||
* - Lessons learned during problem-solving
|
||||
* - Pending tasks or commitments
|
||||
*
|
||||
* Also handles auto-updating USER.md with discovered preferences.
|
||||
*
|
||||
* Reference: ZCLAW_AGENT_INTELLIGENCE_EVOLUTION.md §6.2.2
|
||||
*/
|
||||
|
||||
import { getMemoryManager, type MemoryType } from './agent-memory';
|
||||
import { getAgentIdentityManager } from './agent-identity';
|
||||
|
||||
// === Types ===
|
||||
|
||||
export interface ExtractedItem {
|
||||
content: string;
|
||||
type: MemoryType;
|
||||
importance: number;
|
||||
tags: string[];
|
||||
}
|
||||
|
||||
export interface ExtractionResult {
|
||||
items: ExtractedItem[];
|
||||
saved: number;
|
||||
skipped: number;
|
||||
userProfileUpdated: boolean;
|
||||
}
|
||||
|
||||
export interface ConversationMessage {
|
||||
role: string;
|
||||
content: string;
|
||||
}
|
||||
|
||||
// === Extraction Prompt ===
|
||||
|
||||
const EXTRACTION_PROMPT = `请从以下对话中提取值得长期记住的信息。
|
||||
|
||||
只提取以下类型：
|
||||
- fact: 用户告知的事实（如"我的公司叫 XXX"、"我在做 YYY 项目"）
|
||||
- preference: 用户的偏好（如"我喜欢简洁的回答"、"请用中文"）
|
||||
- lesson: 本次对话的经验教训（如"调用 API 前需要先验证 token"）
|
||||
- task: 未完成的任务或承诺（如"下次帮我检查 XXX"）
|
||||
|
||||
评估规则：
|
||||
- importance 1-3: 临时性、不太重要的信息
|
||||
- importance 4-6: 有一定参考价值的信息
|
||||
- importance 7-9: 重要的持久信息
|
||||
- importance 10: 极其关键的信息
|
||||
|
||||
输出**纯 JSON 数组**，每项包含 content, type, importance, tags[]。
|
||||
如果没有值得记忆的内容，返回空数组 []。
|
||||
不要输出任何其他内容，只输出 JSON。
|
||||
|
||||
对话内容：
|
||||
`;
|
||||
|
||||
// === Memory Extractor ===
|
||||
|
||||
export class MemoryExtractor {
|
||||
private minMessagesForExtraction = 4;
|
||||
private extractionCooldownMs = 30_000; // 30 seconds between extractions
|
||||
private lastExtractionTime = 0;
|
||||
|
||||
/**
|
||||
* Extract memories from a conversation using rule-based heuristics.
|
||||
* This is the Phase 1 approach — no LLM call needed.
|
||||
* Phase 2 will add LLM-based extraction using EXTRACTION_PROMPT.
|
||||
*/
|
||||
async extractFromConversation(
|
||||
messages: ConversationMessage[],
|
||||
agentId: string,
|
||||
conversationId?: string
|
||||
): Promise<ExtractionResult> {
|
||||
// Cooldown check
|
||||
if (Date.now() - this.lastExtractionTime < this.extractionCooldownMs) {
|
||||
return { items: [], saved: 0, skipped: 0, userProfileUpdated: false };
|
||||
}
|
||||
|
||||
// Minimum message threshold
|
||||
const chatMessages = messages.filter(m => m.role === 'user' || m.role === 'assistant');
|
||||
if (chatMessages.length < this.minMessagesForExtraction) {
|
||||
return { items: [], saved: 0, skipped: 0, userProfileUpdated: false };
|
||||
}
|
||||
|
||||
this.lastExtractionTime = Date.now();
|
||||
|
||||
// Phase 1: Rule-based extraction (pattern matching)
|
||||
const extracted = this.ruleBasedExtraction(chatMessages);
|
||||
|
||||
// Save to memory
|
||||
const memoryManager = getMemoryManager();
|
||||
let saved = 0;
|
||||
let skipped = 0;
|
||||
|
||||
for (const item of extracted) {
|
||||
try {
|
||||
await memoryManager.save({
|
||||
agentId,
|
||||
content: item.content,
|
||||
type: item.type,
|
||||
importance: item.importance,
|
||||
source: 'auto',
|
||||
tags: item.tags,
|
||||
conversationId,
|
||||
});
|
||||
saved++;
|
||||
} catch {
|
||||
skipped++;
|
||||
}
|
||||
}
|
||||
|
||||
// Auto-update USER.md with preferences
|
||||
let userProfileUpdated = false;
|
||||
const preferences = extracted.filter(e => e.type === 'preference' && e.importance >= 5);
|
||||
if (preferences.length > 0) {
|
||||
try {
|
||||
const identityManager = getAgentIdentityManager();
|
||||
const prefSummary = preferences.map(p => `- ${p.content}`).join('\n');
|
||||
identityManager.appendToUserProfile(agentId, `### 自动发现的偏好 (${new Date().toLocaleDateString('zh-CN')})\n${prefSummary}`);
|
||||
userProfileUpdated = true;
|
||||
} catch (err) {
|
||||
console.warn('[MemoryExtractor] Failed to update USER.md:', err);
|
||||
}
|
||||
}
|
||||
|
||||
if (saved > 0) {
|
||||
console.log(`[MemoryExtractor] Extracted ${saved} memories from conversation (${skipped} skipped)`);
|
||||
}
|
||||
|
||||
return { items: extracted, saved, skipped, userProfileUpdated };
|
||||
}
|
||||
|
||||
/**
|
||||
* Phase 1: Rule-based extraction using pattern matching.
|
||||
* Extracts common patterns from user messages.
|
||||
*/
|
||||
private ruleBasedExtraction(messages: ConversationMessage[]): ExtractedItem[] {
|
||||
const items: ExtractedItem[] = [];
|
||||
const userMessages = messages.filter(m => m.role === 'user').map(m => m.content);
|
||||
|
||||
for (const msg of userMessages) {
|
||||
// Fact patterns
|
||||
this.extractFacts(msg, items);
|
||||
// Preference patterns
|
||||
this.extractPreferences(msg, items);
|
||||
// Task patterns
|
||||
this.extractTasks(msg, items);
|
||||
}
|
||||
|
||||
// Lesson extraction from assistant messages (error corrections, solutions)
|
||||
const assistantMessages = messages.filter(m => m.role === 'assistant').map(m => m.content);
|
||||
this.extractLessons(userMessages, assistantMessages, items);
|
||||
|
||||
return items;
|
||||
}
|
||||
|
||||
private extractFacts(msg: string, items: ExtractedItem[]): void {
|
||||
// "我的/我们的 X 是/叫 Y" patterns
|
||||
const factPatterns = [
|
||||
/我(?:的|们的|们)(\S{1,20})(?:是|叫|名叫|名字是)(.{2,50})/g,
|
||||
/(?:公司|团队|项目|产品)(?:名|名称)?(?:是|叫)(.{2,30})/g,
|
||||
/我(?:在|正在)(?:做|开发|使用|学习)(.{2,40})/g,
|
||||
/我(?:是|做)(.{2,30})(?:的|工作)/g,
|
||||
];
|
||||
|
||||
for (const pattern of factPatterns) {
|
||||
const matches = msg.matchAll(pattern);
|
||||
for (const match of matches) {
|
||||
const content = match[0].trim();
|
||||
if (content.length > 5 && content.length < 100) {
|
||||
items.push({
|
||||
content,
|
||||
type: 'fact',
|
||||
importance: 6,
|
||||
tags: ['auto-extracted'],
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
private extractPreferences(msg: string, items: ExtractedItem[]): void {
|
||||
const prefPatterns = [
|
||||
/(?:我喜欢|我偏好|我习惯|请用|请使用|默认用|我更愿意)(.{2,50})/g,
|
||||
/(?:不要|别|不用)(.{2,30})(?:了|吧)?/g,
|
||||
/(?:以后|下次|每次)(?:都)?(.{2,40})/g,
|
||||
/(?:用中文|用英文|简洁|详细|简短)(?:一点|回复|回答)?/g,
|
||||
];
|
||||
|
||||
for (const pattern of prefPatterns) {
|
||||
const matches = msg.matchAll(pattern);
|
||||
for (const match of matches) {
|
||||
const content = match[0].trim();
|
||||
if (content.length > 3 && content.length < 80) {
|
||||
items.push({
|
||||
content: `用户偏好: ${content}`,
|
||||
type: 'preference',
|
||||
importance: 5,
|
||||
tags: ['auto-extracted', 'preference'],
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
private extractTasks(msg: string, items: ExtractedItem[]): void {
|
||||
const taskPatterns = [
|
||||
/(?:帮我|帮忙|记得|别忘了|下次|以后|待办)(.{5,60})/g,
|
||||
/(?:TODO|todo|FIXME|fixme)[:\s]*(.{5,60})/g,
|
||||
];
|
||||
|
||||
for (const pattern of taskPatterns) {
|
||||
const matches = msg.matchAll(pattern);
|
||||
for (const match of matches) {
|
||||
const content = match[0].trim();
|
||||
if (content.length > 5 && content.length < 100) {
|
||||
items.push({
|
||||
content,
|
||||
type: 'task',
|
||||
importance: 7,
|
||||
tags: ['auto-extracted', 'task'],
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
private extractLessons(
|
||||
_userMessages: string[],
|
||||
assistantMessages: string[],
|
||||
items: ExtractedItem[]
|
||||
): void {
|
||||
// Look for error resolution patterns in assistant messages
|
||||
for (const msg of assistantMessages) {
|
||||
// "问题是/原因是/根因是" patterns
|
||||
const lessonPatterns = [
|
||||
/(?:问题是|原因是|根因是|解决方法是|关键是)(.{10,100})/g,
|
||||
/(?:需要注意|要注意|注意事项)[：:](.{10,80})/g,
|
||||
];
|
||||
|
||||
for (const pattern of lessonPatterns) {
|
||||
const matches = msg.matchAll(pattern);
|
||||
for (const match of matches) {
|
||||
const content = match[0].trim();
|
||||
if (content.length > 10 && content.length < 150) {
|
||||
items.push({
|
||||
content,
|
||||
type: 'lesson',
|
||||
importance: 6,
|
||||
tags: ['auto-extracted', 'lesson'],
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Build the LLM extraction prompt for a conversation.
|
||||
* For Phase 2: send this to LLM and parse the JSON response.
|
||||
*/
|
||||
buildExtractionPrompt(messages: ConversationMessage[]): string {
|
||||
const conversationText = messages
|
||||
.filter(m => m.role === 'user' || m.role === 'assistant')
|
||||
.map(m => `[${m.role === 'user' ? '用户' : '助手'}]: ${m.content}`)
|
||||
.join('\n\n');
|
||||
|
||||
return EXTRACTION_PROMPT + conversationText;
|
||||
}
|
||||
|
||||
/**
|
||||
* Parse LLM extraction response.
|
||||
* For Phase 2: parse the JSON array from LLM response.
|
||||
*/
|
||||
parseExtractionResponse(response: string): ExtractedItem[] {
|
||||
try {
|
||||
// Find JSON array in response
|
||||
const jsonMatch = response.match(/\[[\s\S]*\]/);
|
||||
if (!jsonMatch) return [];
|
||||
|
||||
const parsed = JSON.parse(jsonMatch[0]);
|
||||
if (!Array.isArray(parsed)) return [];
|
||||
|
||||
return parsed
|
||||
.filter((item: Record<string, unknown>) =>
|
||||
item.content && item.type && item.importance !== undefined
|
||||
)
|
||||
.map((item: Record<string, unknown>) => ({
|
||||
content: String(item.content),
|
||||
type: item.type as MemoryType,
|
||||
importance: Math.max(1, Math.min(10, Number(item.importance))),
|
||||
tags: Array.isArray(item.tags) ? item.tags.map(String) : [],
|
||||
}));
|
||||
} catch {
|
||||
console.warn('[MemoryExtractor] Failed to parse LLM extraction response');
|
||||
return [];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// === Singleton ===
|
||||
|
||||
let _instance: MemoryExtractor | null = null;
|
||||
|
||||
export function getMemoryExtractor(): MemoryExtractor {
|
||||
if (!_instance) {
|
||||
_instance = new MemoryExtractor();
|
||||
}
|
||||
return _instance;
|
||||
}
|
||||
|
||||
export function resetMemoryExtractor(): void {
|
||||
_instance = null;
|
||||
}
|
||||
Reference in New Issue
Block a user