feat: 完成 C 群組 — 記憶子系統

- WorkingMemoryService:session 對話上下文緩衝,token 溢位時優先裁切最舊的低情緒段落
- JobQueue 抽象層(now/schedule(at)/every(cron))與 InProcessJobQueue 實作(node-cron),
  R-6 會換成 BullMQ 但呼叫端介面不變
- MemoryConsolidationService:睡眠固化,只有高情緒或重複提及的內容才寫入 EpisodicMemory
- ForgettingSweepService:低情緒且長期未提取的記憶逐次降權、weight 過低後刪除,
  已掛上每小時一次的 JobQueue.every 排程
- KeywordMemoryRetriever:關鍵字+時間近因+情緒權重排序,提取命中即更新提取次數/時間
- apps/api 新增 /memory/* 端點作為引擎驗證介面(H 群組會決定併入正式對話管線後的去留)
- scripts/smoke/C.mjs:端到端驗證溢位裁切、只寫工作記憶、固化篩選、檢索排序、
  提取即改寫、schedule(at) 準時觸發、遺忘衰減與刪除

npm run restart && npm run smoke -- C 皆通過(C-V),A/B 群組冒煙測試無回歸。
This commit is contained in:
Jeffery
2026-08-13 09:54:29 +08:00
parent f44200f543
commit e6f25008d1
14 changed files with 580 additions and 8 deletions
@@ -0,0 +1,61 @@
import { Injectable } from "@nestjs/common";
import type { EmotionTag } from "@kokorone/shared";
import { HIGH_EMOTION_THRESHOLD } from "./constants.js";
export interface WorkingMemoryEntry {
role: "user" | "character";
content: string;
timestamp: Date;
emotionTag?: EmotionTag;
emotionIntensity?: number; // 0~1,未標記情緒的訊息(如快速通道問候)可省略
}
const DEFAULT_TOKEN_LIMIT = 200;
// 粗略估算:Mock 階段不需要真實 tokenizer,先以字元數/2 近似。
function estimateTokens(text: string): number {
return Math.ceil(text.length / 2);
}
function isHighEmotion(entry: WorkingMemoryEntry): boolean {
return (entry.emotionIntensity ?? 0) >= HIGH_EMOTION_THRESHOLD;
}
// C-1 工作記憶:session 對話上下文緩衝,只存在於行程記憶體中(如同海馬迴暫存),
// 不落地到任何長期記憶表——長期寫入只能由 C-4 睡眠固化觸發。
@Injectable()
export class WorkingMemoryService {
private readonly sessions = new Map<string, WorkingMemoryEntry[]>();
private readonly tokenLimit = DEFAULT_TOKEN_LIMIT;
append(sessionId: string, entry: WorkingMemoryEntry): void {
const entries = this.sessions.get(sessionId) ?? [];
entries.push(entry);
this.sessions.set(sessionId, this.evictOverflow(entries));
}
getContext(sessionId: string): WorkingMemoryEntry[] {
return [...(this.sessions.get(sessionId) ?? [])];
}
clear(sessionId: string): void {
this.sessions.delete(sessionId);
}
private evictOverflow(entries: WorkingMemoryEntry[]): WorkingMemoryEntry[] {
let result = entries;
let totalTokens = result.reduce((sum, e) => sum + estimateTokens(e.content), 0);
while (totalTokens > this.tokenLimit && result.length > 0) {
// 最舊的低情緒段落先被裁掉;若全部都是高情緒,最後才犧牲最舊的一則以確保不超出上限。
let evictIndex = result.findIndex((e) => !isHighEmotion(e));
if (evictIndex === -1) {
evictIndex = 0;
}
const [evicted] = result.splice(evictIndex, 1);
totalTokens -= estimateTokens(evicted.content);
}
return result;
}
}