@memory-river/core 0.2.2 → 0.2.3

package/dist/distill/concentrator-adapter.d.ts

@@ -88,6 +88,7 @@ export declare function findProbeTextMismatchDetail(summarizePairs: ComparableTr
 export declare function logProbeTextMismatchDetail(detail: ProbeTextMismatchDetail): void;
 export declare function buildDualTrackPrompt(conversationLog: string, capsuleLanguage?: string): string;
 export declare function buildGeneralConversationPrompt(conversationLog: string, capsuleLanguage?: string): string;
+export declare function buildImportanceScoringPrompt(noteTexts: string[]): string;
 export declare function buildSimplePrompt(conversationLog: string, capsuleLanguage?: string): string;
 export type CodexReasoningEffort = 'none' | 'low' | 'medium' | 'high' | 'xhigh' | 'max';
 export interface CodexCliConfig {
@@ -129,6 +130,7 @@ export interface ConcentratorConfig {
     transcriptArchive: TranscriptArchive;
     sessionSummaryDir: string;
     llm?: LlmClient;
+    importanceScorer?: (prompt: string) => Promise<string>;
 }
 export interface CapsuleOutput {
     messages: ContextMessage[];
@@ -142,6 +144,7 @@ export declare class ConcentratorAdapter implements LlmClient {
     private statsStore?;
     private transcriptArchive;
     private llm?;
+    private importanceScorer?;
     private readonly MAX_CONTEXT_WINDOW;
     private static readonly WATERLINE_CODE;
     private static readonly WATERLINE_DEFAULT;
@@ -153,6 +156,7 @@ export declare class ConcentratorAdapter implements LlmClient {
      */
     private detectConversationMode;
     buildFallbackCapsule(messages: ContextMessage[]): string;
+    private scoreNoteImportance;
     concentrate(rawMessages: ContextMessage[], dryRun?: boolean, force?: boolean, context?: ConcentrateContext): Promise<CapsuleOutput>;
     estimateTokens(messages: ContextMessage[]): number;
     estimateTokenBreakdown(messages: ContextMessage[]): {

package/dist/distill/concentrator-adapter.js

@@ -796,7 +796,7 @@ ${conversationLog}
 【時間正規化】對話每個 turn 前綴有 [at=<ISO>] 絕對時間戳。當事實含「昨天/今天/明天/上週/下個月/剛才/之後」等相對時間，必須以該事實所在 turn 的 at 為錨，在 text 內寫出絕對日期（例：「在 2023-05-07（原文稱昨天）參加了…」），並填 when 欄位。若 at=unknown 或語意不足以唯一解析，保留原文相對詞、不可猜，when.precision 與 when.source 設為 "unknown"。明確日期、月份、年份、區間、期限、事件先後順序都要保留，不要只寫「最近」「之前」。
 
 每筆格式：
-{ "text": "...", "category": "fact|decision|entity|preference|constraint|identity|knowledge|history|business|other", "importance": 0.0-1.0, "tags": ["..."], "subject": "主體（選配）", "predicate": "關係/屬性/動作（選配）", "value": "值，字串/數字/布林（選配）", "unit": "數值單位（選配）", "when": { "start": "YYYY-MM-DD 或 ISO（選配）", "end": "（選配）", "precision": "datetime|date|month|year|range|unknown", "sourceText": "原文時間詞", "source": "explicit|relative_anchored|contextual|unknown", "anchor": "錨定用的 turn ISO 時間" } }
+{ "text": "...", "category": "fact|decision|entity|preference|constraint|identity|knowledge|history|business|other", "tags": ["..."], "subject": "主體（選配）", "predicate": "關係/屬性/動作（選配）", "value": "值，字串/數字/布林（選配）", "unit": "數值單位（選配）", "when": { "start": "YYYY-MM-DD 或 ISO（選配）", "end": "（選配）", "precision": "datetime|date|month|year|range|unknown", "sourceText": "原文時間詞", "source": "explicit|relative_anchored|contextual|unknown", "anchor": "錨定用的 turn ISO 時間" } }
 （subject/predicate/value/unit/entities/when 全為選配：有對應資訊才填，沒有就整個省略，不可填空字串或編造。text 仍是主要可讀、可檢索內容；日期、數值、人名、地名、檔案名、產品名、事件名也必須同時寫進 text，能正規化的日期同時寫入 text 與 when。）
 
 收錄：
@@ -816,7 +816,6 @@ ${conversationLog}
 - 一筆只表達一個可獨立更新的主張，必須自足：寫明主體、內容、作用域，以及必要的時間、狀態、條件或理由；不得使用「這個」「上述方案」「已處理」等脫離原文便無法理解的指涉。
 - 保留足以語意檢索與回查原始 transcript 的專案、元件、人物或事件名稱，但不要複製整段對話、長篇推理、日誌或操作過程（原文可由 rehydrate 取回）。
 - category 選最能代表該主張長期用途者；tags 用少量具辨識力的實體/專案/領域/狀態詞。
-- importance 依「跨時間耐久性 × 未來決策效用」評分，不依篇幅、情緒強度或主題是否熱門；較低分代表耐久性或決策效用較低，但不得因此省略可獨立檢索的具體事實。
 
 CRITICAL INSTRUCTION: Output ONLY valid, raw JSON. Do NOT wrap in markdown code blocks. Start with '{' and end with '}'.
 
@@ -843,7 +842,6 @@ confidence 評分標準（0.0–1.0）：
     {
       "text": "精確的顆粒化記憶（含足夠上下文，獨立可理解；日期/數值/人名要寫進 text）",
       "category": "fact|decision|entity|preference|constraint|identity|knowledge|history|business|other",
-      "importance": 0.0-1.0,
       "tags": [],
       "subject": "主體（選配，省略則整個不要出現）",
       "predicate": "關係/屬性/動作（選配）",
@@ -898,7 +896,7 @@ ${capsuleLanguageInstruction}寫一段「前情提要」（600–900 字），
 【時間正規化】對話每個 turn 前綴有 [at=<ISO>] 絕對時間戳。當事實含「昨天/今天/明天/上週/下個月/剛才/之後」等相對時間，必須以該事實所在 turn 的 at 為錨，在 text 內寫出絕對日期（例：「在 2023-05-07（原文稱昨天）參加了…」），並填 when 欄位。若 at=unknown 或語意不足以唯一解析，保留原文相對詞、不可猜，when.precision 與 when.source 設為 "unknown"。明確日期、月份、年份、區間、期限、事件先後順序都要保留，不要只寫「最近」「之前」。
 
 每筆格式：
-{ "text": "...", "category": "fact|decision|entity|preference|constraint|identity|knowledge|history|business|other", "importance": 0.0-1.0, "tags": ["..."], "subject": "主體（選配）", "predicate": "關係/屬性/動作（選配）", "value": "值，字串/數字/布林（選配）", "unit": "數值單位（選配）", "when": { "start": "YYYY-MM-DD 或 ISO（選配）", "end": "（選配）", "precision": "datetime|date|month|year|range|unknown", "sourceText": "原文時間詞", "source": "explicit|relative_anchored|contextual|unknown", "anchor": "錨定用的 turn ISO 時間" } }
+{ "text": "...", "category": "fact|decision|entity|preference|constraint|identity|knowledge|history|business|other", "tags": ["..."], "subject": "主體（選配）", "predicate": "關係/屬性/動作（選配）", "value": "值，字串/數字/布林（選配）", "unit": "數值單位（選配）", "when": { "start": "YYYY-MM-DD 或 ISO（選配）", "end": "（選配）", "precision": "datetime|date|month|year|range|unknown", "sourceText": "原文時間詞", "source": "explicit|relative_anchored|contextual|unknown", "anchor": "錨定用的 turn ISO 時間" } }
 （subject/predicate/value/unit/entities/when 全為選配：有對應資訊才填，沒有就整個省略，不可填空字串或編造。text 仍是主要可讀、可檢索內容；日期、數值、人名、地名、檔案名、產品名、事件名也必須同時寫進 text，能正規化的日期同時寫入 text 與 when。）
 
 收錄：
@@ -918,7 +916,6 @@ ${capsuleLanguageInstruction}寫一段「前情提要」（600–900 字），
 - 一筆只表達一個可獨立更新的主張，必須自足：寫明主體、內容、作用域，以及必要的時間、狀態、條件或理由；不得使用「這個」「上述方案」「已處理」等脫離原文便無法理解的指涉。
 - 保留足以語意檢索與回查原始 transcript 的專案、元件、人物或事件名稱，但不要複製整段對話、長篇推理、日誌或操作過程（原文可由 rehydrate 取回）。
 - category 選最能代表該主張長期用途者；tags 用少量具辨識力的實體/專案/領域/狀態詞。
-- importance 依「跨時間耐久性 × 未來決策效用」評分，不依篇幅、情緒強度或主題是否熱門；較低分代表耐久性或決策效用較低，但不得因此省略可獨立檢索的具體事實。
 
 CRITICAL INSTRUCTION: Output ONLY valid, raw JSON. Do NOT wrap in markdown code blocks. Start with '{' and end with '}'.
 
@@ -932,7 +929,7 @@ confidence 評分標準（0.0–1.0）：
   "confidence": 0.85,
   "capsule": "任務B的自然語言前情提要...",
   "notes": [
-    { "text": "精確的顆粒化記憶（含足夠上下文，獨立可理解；日期/數值/人名要寫進 text）", "category": "fact|decision|entity|preference|constraint|identity|knowledge|history|business|other", "importance": 0.0, "tags": [], "subject": "（選配，省略則不要出現）", "predicate": "（選配）", "value": "（選配）", "unit": "（選配）", "when": { "start": "YYYY-MM-DD（選配）", "precision": "date", "sourceText": "原文時間詞", "source": "relative_anchored", "anchor": "turn ISO 時間" } }
+    { "text": "精確的顆粒化記憶（含足夠上下文，獨立可理解；日期/數值/人名要寫進 text）", "category": "fact|decision|entity|preference|constraint|identity|knowledge|history|business|other", "tags": [], "subject": "（選配，省略則不要出現）", "predicate": "（選配）", "value": "（選配）", "unit": "（選配）", "when": { "start": "YYYY-MM-DD（選配）", "precision": "date", "sourceText": "原文時間詞", "source": "relative_anchored", "anchor": "turn ISO 時間" } }
   ]
 }
 
@@ -942,6 +939,70 @@ confidence 評分標準（0.0–1.0）：
 - 必須回傳可完整 JSON.parse 的有效 JSON，不可輸出半截 JSON`;
     return isSourceLanguage ? `${sourceLanguageRequirementTop}\n\n${prompt}\n\n${sourceLanguageRequirementBottom}` : prompt;
 }
+export function buildImportanceScoringPrompt(noteTexts) {
+    const numberedNotes = noteTexts
+        .map((text, index) => `${index + 1}. ${text}`)
+        .join('\n');
+    return `非互動模式,立即回答,不要提問、不要使用任何工具、不要讀檔。
+
+下面是即將存進 AI 長期記憶庫的記憶。請為每一則評 importance(0.0-1.0)。
+
+評分標準(照這個定義,不要自行更改):
+importance 依「跨時間耐久性 × 未來決策效用」評分,不依篇幅、情緒強度或主題是否熱門;
+較低分代表耐久性或決策效用較低。
+
+只輸出 JSON,不要任何其他文字、不要 markdown 圍欄。必須包含全部 <${noteTexts.length}> 筆:
+{"r":[{"n":<編號>,"i":<0.0-1.0>}]}
+
+${numberedNotes}`;
+}
+class ImportanceScoreParseError extends Error {
+    scores;
+    constructor(message, scores) {
+        super(message);
+        this.scores = scores;
+        this.name = 'ImportanceScoreParseError';
+    }
+}
+function parseImportanceScores(raw, expectedCount) {
+    const lines = raw.split(/\r?\n/).map((line) => line.trim()).filter(Boolean);
+    let response = null;
+    for (let index = lines.length - 1; index >= 0; index--) {
+        try {
+            const candidate = JSON.parse(lines[index]);
+            if (candidate && typeof candidate === 'object' && Array.isArray(candidate.r)) {
+                response = candidate;
+                break;
+            }
+        }
+        catch {
+            // Codex progress lines and other non-JSON output are expected here.
+        }
+    }
+    if (!response) {
+        throw new ImportanceScoreParseError('importance score response parse failed', new Map());
+    }
+    const scores = new Map();
+    for (const entry of response.r) {
+        if (!entry || typeof entry !== 'object')
+            continue;
+        const item = entry;
+        if (typeof item.n === 'number'
+            && Number.isInteger(item.n)
+            && item.n >= 1
+            && item.n <= expectedCount
+            && typeof item.i === 'number'
+            && Number.isFinite(item.i)) {
+            scores.set(item.n, Math.max(0, Math.min(1, item.i)));
+        }
+    }
+    const missing = Array.from({ length: expectedCount }, (_, index) => index + 1)
+        .filter((number) => !scores.has(number));
+    if (missing.length > 0) {
+        throw new ImportanceScoreParseError(`importance score response missing notes: ${missing.join(',')}`, scores);
+    }
+    return scores;
+}
 // ─── 降級版 Prompt (專供本地小模型使用) ───
 export function buildSimplePrompt(conversationLog, capsuleLanguage = '繁體中文') {
     if (capsuleLanguage === 'source') {
@@ -1396,6 +1457,7 @@ export class ConcentratorAdapter {
     statsStore;
     transcriptArchive;
     llm;
+    importanceScorer;
     MAX_CONTEXT_WINDOW = 200000;
     // ── 動態水位線常數 ──────────────────────────────────────
     static WATERLINE_CODE = 0.40; // 代碼/技術密集：40%
@@ -1422,6 +1484,7 @@ export class ConcentratorAdapter {
         this.statsStore = config.statsStore;
         this.transcriptArchive = config.transcriptArchive;
         this.llm = config.llm;
+        this.importanceScorer = config.importanceScorer;
         this.capsuleBridge = new CapsuleBridge(this.config.inboxPath);
     }
     /**
@@ -1491,6 +1554,41 @@ export class ConcentratorAdapter {
             `## 9. Optional Next Step\n${latestAssistant === '無' ? '無' : latestAssistant}`,
         ].join('\n\n');
     }
+    async scoreNoteImportance(notes, metricContext) {
+        if (notes.length === 0)
+            return;
+        const prompt = buildImportanceScoringPrompt(notes.map((note) => typeof note.text === 'string' ? note.text : ''));
+        let scores = new Map();
+        try {
+            const raw = this.importanceScorer
+                ? await this.importanceScorer(prompt)
+                : await this.callWithFallback(prompt, 'scoreImportance', undefined, { sessionIdentity: metricContext?.sessionIdentity, inputTokens: estimatePromptTokens(prompt) }, 2048);
+            scores = parseImportanceScores(raw, notes.length);
+        }
+        catch (error) {
+            if (error instanceof ImportanceScoreParseError) {
+                scores = error.scores;
+            }
+            const message = String(error?.message ?? error).toLowerCase();
+            const reason = error instanceof ImportanceScoreParseError
+                ? (message.includes('missing') ? 'partial missing' : 'parse failed')
+                : (message.includes('timeout') || message.includes('timed out') || message.includes('abort')
+                    ? 'timeout'
+                    : 'call failed');
+            console.warn(`[ConcentratorAdapter] Importance scoring fallback: ${reason}; missing scores use 0.5`);
+        }
+        notes.forEach((note, index) => {
+            const importance = scores.get(index + 1);
+            if (importance === undefined) {
+                note.importance = 0.5;
+                note.importanceFallback = true;
+            }
+            else {
+                note.importance = importance;
+                delete note.importanceFallback;
+            }
+        });
+    }
     async concentrate(rawMessages, dryRun = false, force = false, context = {}) {
         const messages = rawMessages.map(msg => truncateGiantStrings(msg));
         const currentTokens = estimateTotalTokens(messages);
@@ -1692,7 +1790,9 @@ export class ConcentratorAdapter {
                 const originalTokens = estimatePromptTokens(conversationLog);
                 const summaryTokens = estimatePromptTokens(capsuleText);
                 const compressionRatio = originalTokens / Math.max(1, summaryTokens);
-                const notes = parsedData.notes || [];
+                const notes = Array.isArray(parsedData.notes)
+                    ? parsedData.notes.map((item) => item && typeof item === 'object' ? item : {})
+                    : [];
                 const sourceEntryIds = sourceEntryIdsProbe?.matched ? sourceEntryIdsProbe.sourceEntryIds : [];
                 const sourceEntryRange = sourceEntryIds.length > 0
                     ? {
@@ -1720,6 +1820,9 @@ export class ConcentratorAdapter {
                     console.log(`[ConcentratorAdapter] sourceEntryIds metadata: length=${sourceEntryIds.length} firstEntryId=${sourceEntryRange?.firstEntryId ?? 'none'} lastEntryId=${sourceEntryRange?.lastEntryId ?? 'none'}`);
                     console.log(`[ConcentratorAdapter] Short-term capsule written (health: 30, confidence: ${confidence.toFixed(2)})`);
                 }
+                await this.scoreNoteImportance(notes, {
+                    sessionIdentity: context.sessionIdentity,
+                });
                 const acceptedNoteTexts = [];
                 for (const item of notes) {
                     const text = typeof item.text === 'string' ? item.text.trim() : '';
@@ -1744,6 +1847,7 @@ export class ConcentratorAdapter {
                                 confidence,
                                 firstTimestamp,
                                 lastTimestamp,
+                                ...(item.importanceFallback === true ? { importanceFallback: true } : {}),
                                 ...(sourceEntryIds.length > 0 ? { sourceEntryIds } : {}),
                                 ...(sourceEntryRange ? { sourceEntryRange } : {}),
                                 // Optional structured enrichment — additive, stored as-is (no schema change),
@@ -1770,7 +1874,7 @@ export class ConcentratorAdapter {
                     || rawMessages[0]?.sessionId
                     || 'unknown';
                 const concentratedAt = Date.now();
-                await this.writeSessionSummary({ sessionId, concentratedAt, capsule: capsuleText, notes: parsedData.notes || [], primaryRequest: summary?.primaryRequest || '', pendingTasks: summary?.pendingTasks || '', nextStep: summary?.nextStep || '' });
+                await this.writeSessionSummary({ sessionId, concentratedAt, capsule: capsuleText, notes: notes, primaryRequest: summary?.primaryRequest || '', pendingTasks: summary?.pendingTasks || '', nextStep: summary?.nextStep || '' });
             }
             catch (err) {
                 console.error("[ConcentratorAdapter] Compaction failed:", err);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@memory-river/core",
-  "version": "0.2.2",
+  "version": "0.2.3",
   "description": "Framework-neutral long-term memory engine for Node.js agents.",
   "private": false,
   "license": "Apache-2.0",