@scotthuang/engram 0.6.8 → 0.7.0

package/dist/settle.js

@@ -1,675 +0,0 @@
-/**
- * Memory System Plugin - Settle (沉淀机制)
- *
- * 双层架构（v0.5）：
- *
- * ■ 每日 settle (runSettlement):
- *   1. 结构化短期记忆
- *   2. 更新画像
- *   3. 生成画像摘要
- *   短期文件永久保留，不再归档删除
- *
- * ■ 月度 settle (runMonthlySettle):
- *   1. 精选沉淀：只沉淀 recall-hits 中 hitCount ≥ 2 的短期记忆到向量库
- *   2. 月度遗忘：清理向量库中低 effective_importance 的条目
- */
-import { promises as fs } from "node:fs";
-import { logger } from "./logger.js";
-import { join } from "node:path";
-import { ProfileManager } from "./profile.js";
-/**
- * 获取本地日期字符串 (YYYY-MM-DD)
- * 使用本地时区而非 UTC，避免跨日期问题
- */
-function getLocalDateString() {
-    const now = new Date();
-    const year = now.getFullYear();
-    const month = String(now.getMonth() + 1).padStart(2, "0");
-    const day = String(now.getDate()).padStart(2, "0");
-    return `${year}-${month}-${day}`;
-}
-const STRUCTURED_MARKER = "<!-- STRUCTURED -->";
-const SETTLED_MARKER = "<!-- SETTLED -->";
-/**
- * 清理 LLM 返回的 JSON：去除 markdown 代码块包裹、前后空白等
- * 处理常见格式：```json\n...\n```、```\n...\n```
- */
-function cleanLLMJson(raw) {
-    let s = raw.trim();
-    // 去除 ```json ... ``` 或 ``` ... ``` 包裹
-    s = s.replace(/^```(?:json)?\s*\n?/i, "").replace(/\n?\s*```\s*$/, "");
-    return s.trim();
-}
-/**
- * 步骤 1: 结构化短期记忆
- * 读取 short-term 文件，对未结构化的段落进行 LLM 整理
- * 已结构化的段落（含 STRUCTURED_MARKER）跳过
- */
-export async function structurizeShortTerm(opts, date = getLocalDateString()) {
-    const shortTermDir = join(opts.workspaceDir, "memory-engram", "short-term");
-    await fs.mkdir(shortTermDir, { recursive: true });
-    const filePath = join(shortTermDir, `${date}.md`);
-    try {
-        const raw = await fs.readFile(filePath, "utf-8");
-        logger.info(`[engram:settle] Step1 structurize: reading ${filePath} (${raw.length} chars)`);
-        if (!raw.trim()) {
-            logger.info(`[engram:settle] Step1 structurize: file empty, skipping`);
-            return "No content to structurize.";
-        }
-        // 按 ### 拆分段落
-        const sections = raw.split(/(?=^### )/m);
-        logger.info(`[engram:settle] Step1 structurize: ${sections.length} sections found`);
-        let processedCount = 0;
-        let skippedStructured = 0;
-        let skippedSystem = 0;
-        const outputSections = [];
-        for (const section of sections) {
-            if (!section.trim())
-                continue;
-            // 已标记结构化的直接保留
-            if (section.includes(STRUCTURED_MARKER)) {
-                outputSections.push(section);
-                skippedStructured++;
-                continue;
-            }
-            // 纯系统/compact 记录，标记为已处理
-            if (section.includes("[系统]") && section.includes("Compaction triggered")) {
-                outputSections.push(section.trimEnd() + "\n" + STRUCTURED_MARKER + "\n");
-                skippedSystem++;
-                continue;
-            }
-            // 未结构化 → LLM 整理
-            if (!opts.llmCall) {
-                logger.info(`[engram:settle] Step1 structurize: no llmCall, keeping raw section`);
-                outputSections.push(section);
-                continue;
-            }
-            const systemPrompt = `你是一个记忆整理助手。请将以下原始对话记录整理成结构化格式。
-
-规则：
-1. 按时间倒序排列
-2. 每条记录格式：### HH:MM [分类标签]\n摘要内容（一句话）
-3. 过滤掉无意义的闲聊（打招呼、HEARTBEAT_OK、确认回复、"好的"、"嗯"等）
-4. 保留所有有价值的信息（决策、偏好、事件、数字、人名、地点等）
-5. 分类标签可选值：饮食、工作、家庭、技术、决策、健康、购物、出行、随聊
-
-只输出整理后的内容，不要解释。`;
-            try {
-                logger.info(`[engram:settle] Step1: calling LLM to structurize section ${processedCount + skippedStructured + skippedSystem + 1} (${section.length} chars)`);
-                const result = await opts.llmCall(section, systemPrompt);
-                logger.info(`[engram:settle] Step1: LLM returned ${result.length} chars for section`);
-                outputSections.push(result.trimEnd() + "\n" + STRUCTURED_MARKER + "\n");
-                processedCount++;
-            }
-            catch (llmErr) {
-                logger.error(`[engram:settle] Step1: LLM structurize failed for section: ${llmErr}`);
-                // LLM 调用失败，保留原文
-                outputSections.push(section);
-            }
-        }
-        // 写回文件
-        await fs.writeFile(filePath, outputSections.join("\n"), "utf-8");
-        logger.info(`[engram:settle] Step1: done. processed=${processedCount} skippedStructured=${skippedStructured} skippedSystem=${skippedSystem}`);
-        return `Structurized short-term/${date}.md: ${processedCount} sections processed.`;
-    }
-    catch (err) {
-        logger.error(`[engram:settle] Step1 structurizeShortTerm failed: ${err}`);
-        return `Skipped: short-term/${date}.md not found or error: ${err}`;
-    }
-}
-/**
- * 步骤 2: 筛选有价值条目 + 向量化
- * 只处理已结构化(STRUCTURED)但未沉淀(SETTLED)的段落，避免重复 LLM 调用
- */
-export async function extractAndVectorize(opts, date = new Date().toISOString().split("T")[0]) {
-    const filePath = join(opts.workspaceDir, "memory-engram", "short-term", `${date}.md`);
-    logger.info(`[engram:settle] Step2 extractAndVectorize: date=${date}, file=${filePath}`);
-    try {
-        const content = await fs.readFile(filePath, "utf-8");
-        if (!content.trim()) {
-            logger.info(`[engram:settle] Step2: file empty`);
-            return "No content.";
-        }
-        if (!opts.llmCall || !opts.vectorStore) {
-            logger.info(`[engram:settle] Step2: missing llmCall=${!!opts.llmCall} vectorStore=${!!opts.vectorStore}`);
-            return "No LLM/vector configured.";
-        }
-        // 拆分段落，区分：未沉淀的新段落 vs 已沉淀的旧段落
-        const allSections = content.split(/(?=^### )/m);
-        const unsettledSections = [];
-        const unsettledIndices = [];
-        let alreadySettled = 0;
-        for (let i = 0; i < allSections.length; i++) {
-            const s = allSections[i];
-            if (s.includes(SETTLED_MARKER)) {
-                // 已沉淀，跳过
-                alreadySettled++;
-            }
-            else if (s.includes(STRUCTURED_MARKER)) {
-                // 已结构化但未沉淀 → 需要处理
-                unsettledSections.push(s);
-                unsettledIndices.push(i);
-            }
-            // 既不 STRUCTURED 也不 SETTLED 的段落 → 等 Step 1 结构化后再处理
-        }
-        logger.info(`[engram:settle] Step2: ${allSections.length} total, ${unsettledSections.length} unsettled, ${alreadySettled} already settled`);
-        if (unsettledSections.length === 0) {
-            logger.info(`[engram:settle] Step2: no new content to extract (all settled)`);
-            return "No new content to extract (all settled).";
-        }
-        const unsettledContent = unsettledSections.join("\n");
-        const systemPrompt = `你是一个信息筛选助手。从以下结构化记忆中筛选出值得长期保留的条目。
-
-保留标准：
-- 包含用户偏好、决策、重要事件、具体信息（地点、金额、人名）
-- 包含用户行为模式和生活习惯（作息时间、运动习惯、饮食规律等）
-- 去掉纯闲聊、临时信息
-
-对每条保留的条目，输出 JSON 数组，每条包含：
-- text: 条目摘要
-- category: 分类
-- importance: 重要度评分 0-1，标准如下：
-  - 0.9-1.0: 关键决策、重要事件（搬家、换工作、关系变化）
-  - 0.7-0.8: 明确偏好、项目里程碑、具体计划
-  - 0.4-0.6: 一般性讨论、技术问答
-  - 0.1-0.3: 日常闲聊、低价值内容
-
-示例输出：
-[{"text": "用户决定搬到深圳", "category": "生活", "importance": 0.9}, {"text": "用户在调试BM25分词", "category": "技术", "importance": 0.5}]
-
-只输出 JSON，不要其他内容。`;
-        logger.info(`[engram:settle] Step2: calling LLM for extraction (${unsettledContent.length} chars)`);
-        const result = await opts.llmCall(unsettledContent, systemPrompt);
-        logger.info(`[engram:settle] Step2: LLM returned ${result.length} chars`);
-        let items;
-        try {
-            const cleaned = cleanLLMJson(result);
-            items = JSON.parse(cleaned);
-            logger.info(`[engram:settle] Step2: parsed ${items.length} candidate items`);
-        }
-        catch (parseErr) {
-            logger.error(`[engram:settle] Step2: JSON parse failed: ${parseErr}, raw="${result.slice(0, 200)}"`);
-            return `Skipped: JSON parse failed`;
-        }
-        let stored = 0;
-        let duplicates = 0;
-        let skippedInvalid = 0;
-        for (const item of items) {
-            if (item.text && item.category) {
-                const imp = typeof item.importance === "number" ? item.importance : 0.5;
-                logger.info(`[engram:settle] Step2: vectorizing [${item.category}] importance=${imp.toFixed(2)} "${item.text.slice(0, 60)}"`);
-                const storeResult = await opts.vectorStore(item.text, item.category, imp);
-                if (storeResult === "created")
-                    stored++;
-                else
-                    duplicates++;
-            }
-            else {
-                skippedInvalid++;
-                logger.info(`[engram:settle] Step2: skipping invalid item: ${JSON.stringify(item).slice(0, 100)}`);
-            }
-        }
-        // 向量化成功后，给已处理的段落打上 SETTLED 标记，写回文件
-        for (const idx of unsettledIndices) {
-            allSections[idx] = allSections[idx].replace(STRUCTURED_MARKER, STRUCTURED_MARKER + "\n" + SETTLED_MARKER);
-        }
-        await fs.writeFile(filePath, allSections.join(""), "utf-8");
-        logger.info(`[engram:settle] Step2: marked ${unsettledIndices.length} sections as SETTLED`);
-        logger.info(`[engram:settle] Step2: done. stored=${stored} duplicates=${duplicates} invalid=${skippedInvalid}`);
-        return `Extracted ${stored} items (${duplicates} dup, ${unsettledIndices.length} sections settled).`;
-    }
-    catch (err) {
-        logger.error(`[engram:settle] Step2 extractAndVectorize failed: ${err}`);
-        return `Skipped: ${err}`;
-    }
-}
-/**
- * 步骤 3: 更新画像
- * 只处理已结构化(STRUCTURED)但未沉淀(SETTLED)的段落
- */
-export async function updateProfile(opts, date = new Date().toISOString().split("T")[0]) {
-    const filePath = join(opts.workspaceDir, "memory-engram", "short-term", `${date}.md`);
-    logger.info(`[engram:settle] Step3 updateProfile: date=${date}`);
-    try {
-        const content = await fs.readFile(filePath, "utf-8");
-        if (!content.trim()) {
-            logger.info(`[engram:settle] Step3: file empty`);
-            return "No content.";
-        }
-        if (!opts.llmCall) {
-            logger.info(`[engram:settle] Step3: no llmCall`);
-            return "No LLM configured.";
-        }
-        // 只取已结构化但未沉淀的段落（避免重复分析）
-        const unsettledContent = content
-            .split(/(?=^### )/m)
-            .filter(s => s.includes(STRUCTURED_MARKER) && !s.includes(SETTLED_MARKER))
-            .join("\n");
-        if (!unsettledContent.trim()) {
-            logger.info(`[engram:settle] Step3: no new content for profile (all settled)`);
-            return "No new content for profile.";
-        }
-        const profileManager = new ProfileManager(opts.workspaceDir);
-        const profile = await profileManager.load();
-        const existingTagCount = Object.values(profile.tags).reduce((sum, tags) => sum + tags.length, 0);
-        logger.info(`[engram:settle] Step3: loaded profile with ${existingTagCount} existing tags across ${Object.keys(profile.tags).length} dimensions`);
-        const systemPrompt = `你是一个用户画像分析助手。从以下记忆中抽取用户标签，更新画像。
-
-当前画像：
-${JSON.stringify(profile.tags, null, 2)}
-
-请输出 JSON：
-{
-  "added": [{"dimension": "分类", "value": "标签"}],
-  "removed": [{"dimension": "分类", "value": "标签"}],
-  "reason": "简要说明"
-}
-
-规则：
-- 只添加有充分依据的标签（至少 2 条记忆佐证）
-- 如果旧标签与新信息矛盾，放入 removed
-- 不要重复添加已有标签
-
-只输出 JSON。`;
-        logger.info(`[engram:settle] Step3: calling LLM for profile update (${unsettledContent.length} chars)`);
-        const result = await opts.llmCall(unsettledContent, systemPrompt);
-        logger.info(`[engram:settle] Step3: LLM returned ${result.length} chars`);
-        let changes;
-        try {
-            const cleaned = cleanLLMJson(result);
-            changes = JSON.parse(cleaned);
-        }
-        catch (parseErr) {
-            logger.error(`[engram:settle] Step3: JSON parse failed: ${parseErr}, raw="${result.slice(0, 200)}"`);
-            return `Skipped: JSON parse failed`;
-        }
-        let added = 0;
-        let removed = 0;
-        if (changes.added) {
-            for (const tag of changes.added) {
-                logger.info(`[engram:settle] Step3: adding tag [${tag.dimension}] "${tag.value}"`);
-                profileManager.addTag(profile, tag.dimension, tag.value);
-                added++;
-            }
-        }
-        if (changes.removed) {
-            for (const tag of changes.removed) {
-                if (profile.tags[tag.dimension]) {
-                    logger.info(`[engram:settle] Step3: removing tag [${tag.dimension}] "${tag.value}"`);
-                    profile.tags[tag.dimension] = profile.tags[tag.dimension].filter(t => t.value !== tag.value);
-                    removed++;
-                }
-            }
-        }
-        if (changes.reason) {
-            logger.info(`[engram:settle] Step3: reason="${changes.reason}"`);
-        }
-        // 衰减未更新的标签
-        profileManager.decayTags(profile, 0.98);
-        await profileManager.save(profile);
-        logger.info(`[engram:settle] Step3: done. added=${added} removed=${removed}`);
-        return `Profile updated: +${added} -${removed}`;
-    }
-    catch (err) {
-        logger.error(`[engram:settle] Step3 updateProfile failed: ${err}`);
-        return `Skipped: ${err}`;
-    }
-}
-/**
- * 步骤 4: 生成画像摘要
- */
-export async function generateProfileSummary(opts) {
-    logger.info(`[engram:settle] Step4 generateProfileSummary: start`);
-    if (!opts.llmCall) {
-        logger.info(`[engram:settle] Step4: no llmCall configured`);
-        return "No LLM configured.";
-    }
-    const profileManager = new ProfileManager(opts.workspaceDir);
-    const profile = await profileManager.load();
-    const allTags = Object.entries(profile.tags)
-        .map(([dim, tags]) => `${dim}: ${tags.map(t => t.value).join(", ")}`)
-        .join("\n");
-    if (!allTags) {
-        logger.info(`[engram:settle] Step4: no tags to summarize`);
-        return "No tags to summarize.";
-    }
-    const systemPrompt = `将以下用户画像标签压缩为一段 100 字以内的中文摘要，用于 AI 检索时快速理解用户特征。
-
-标签：
-${allTags}
-
-只输出摘要文本，不要其他内容。`;
-    try {
-        logger.info(`[engram:settle] Step4: calling LLM for summary (${allTags.length} chars of tags)`);
-        const summary = await opts.llmCall(allTags, systemPrompt);
-        logger.info(`[engram:settle] Step4: LLM returned summary (${summary.length} chars)`);
-        profile.summary = summary.trim();
-        // 更新 coreTags：取每个维度置信度最高的标签
-        const coreTags = [];
-        for (const tags of Object.values(profile.tags)) {
-            const sorted = [...tags].sort((a, b) => b.confidence - a.confidence);
-            if (sorted.length > 0) {
-                coreTags.push(sorted[0].value);
-            }
-        }
-        profile.coreTags = coreTags.slice(0, 10);
-        await profileManager.save(profile);
-        logger.info(`[engram:settle] Step4: done. coreTags=[${profile.coreTags.join(", ")}]`);
-        return `Summary generated: "${summary.trim()}"`;
-    }
-    catch (err) {
-        logger.error(`[engram:settle] Step4 generateProfileSummary failed: ${err}`);
-        return `Failed: ${err}`;
-    }
-}
-/**
- * 步骤 5: 归档清理（支持动态保留）
- * - 未被 recall 命中过的文件：shortTermDays 后归档
- * - 被 recall 命中过的文件：shortTermDays * 3 后归档（延长保留）
- *
- * @param recallHits recall-hits.json 内容，由外部传入（可选）
- */
-export async function archiveShortTerm(opts, recallHits) {
-    const shortTermDir = join(opts.workspaceDir, "memory-engram", "short-term");
-    const coldStorageDir = join(opts.workspaceDir, "memory-engram", "cold-storage");
-    const baseDays = opts.config.shortTermDays;
-    const extendedDays = baseDays * 3;
-    const baseCutoff = new Date();
-    baseCutoff.setDate(baseCutoff.getDate() - baseDays);
-    const extendedCutoff = new Date();
-    extendedCutoff.setDate(extendedCutoff.getDate() - extendedDays);
-    logger.info(`[engram:settle] Step5 archiveShortTerm: baseDays=${baseDays} extendedDays=${extendedDays} baseCutoff=${baseCutoff.toISOString().split("T")[0]} extendedCutoff=${extendedCutoff.toISOString().split("T")[0]}`);
-    // 如果没有传入 recallHits，尝试从文件读取
-    const hits = recallHits ?? await loadRecallHitsFile(opts.workspaceDir);
-    const hitFileCount = Object.keys(hits).length;
-    if (hitFileCount > 0) {
-        logger.info(`[engram:settle] Step5: ${hitFileCount} files have recall hits, will use extended retention`);
-    }
-    try {
-        await fs.mkdir(coldStorageDir, { recursive: true });
-        const files = await fs.readdir(shortTermDir);
-        const mdFiles = files.filter(f => f.endsWith(".md"));
-        logger.info(`[engram:settle] Step5: found ${mdFiles.length} md files in short-term`);
-        let archived = 0;
-        let retained = 0;
-        for (const file of files) {
-            if (!file.endsWith(".md"))
-                continue;
-            const filePath = join(shortTermDir, file);
-            const stat = await fs.stat(filePath);
-            // 判断该文件是否被 recall 命中过
-            const isHit = !!hits[file];
-            const cutoff = isHit ? extendedCutoff : baseCutoff;
-            if (stat.mtime < cutoff) {
-                // 按月归档
-                const monthMatch = file.match(/(\d{4}-\d{2})/);
-                const month = monthMatch ? monthMatch[1] : "unknown";
-                const destDir = join(coldStorageDir, month);
-                await fs.mkdir(destDir, { recursive: true });
-                await fs.rename(filePath, join(destDir, file));
-                logger.info(`[engram:settle] Step5: archived ${file} → cold-storage/${month}/ (hit=${isHit})`);
-                archived++;
-                // 归档后从 recall-hits 中清除
-                if (isHit) {
-                    delete hits[file];
-                }
-            }
-            else if (isHit && stat.mtime < baseCutoff) {
-                // 超过基础天数但未超过扩展天数，被召回命中所以动态保留
-                retained++;
-                logger.info(`[engram:settle] Step5: retained ${file} (hit=true, extended retention)`);
-            }
-        }
-        // 写回更新后的 recall-hits（清除已归档的条目）
-        if (hitFileCount > 0) {
-            await saveRecallHitsFile(opts.workspaceDir, hits);
-        }
-        logger.info(`[engram:settle] Step5: done. archived=${archived} retained=${retained} files`);
-        return `Archived ${archived} files to cold-storage${retained > 0 ? `, retained ${retained} hit files` : ""}.`;
-    }
-    catch (err) {
-        logger.error(`[engram:settle] Step5 archiveShortTerm failed: ${err}`);
-        return `Archive skipped: ${err}`;
-    }
-}
-/**
- * 读取 recall-hits.json（settle 模块内部辅助）
- */
-async function loadRecallHitsFile(workspaceDir) {
-    try {
-        const filePath = join(workspaceDir, "memory-engram", "recall-hits.json");
-        const raw = await fs.readFile(filePath, "utf-8");
-        return JSON.parse(raw);
-    }
-    catch {
-        return {};
-    }
-}
-/**
- * 写入 recall-hits.json
- */
-async function saveRecallHitsFile(workspaceDir, hits) {
-    const filePath = join(workspaceDir, "memory-engram", "recall-hits.json");
-    await fs.writeFile(filePath, JSON.stringify(hits, null, 2), "utf-8");
-}
-/**
- * 步骤 6: 清理过期长期记忆（遗忘机制）
- * 使用 effective_importance = importance × decay(age) × log2(1 + accessCount)
- * 低于阈值的记忆从向量库中删除
- */
-export async function pruneStaleMemories(opts) {
-    logger.info(`[engram:settle] Step6 pruneStaleMemories: start`);
-    if (!opts.vectorPrune) {
-        logger.info(`[engram:settle] Step6: no vectorPrune callback, skipping`);
-        return "No vector prune configured.";
-    }
-    try {
-        logger.info(`[engram:settle] Step6: calling vectorPrune (threshold=0.05, maxPrune=50)`);
-        const pruned = await opts.vectorPrune(0.05, 50);
-        logger.info(`[engram:settle] Step6: done. pruned=${pruned} stale memories`);
-        return `Pruned ${pruned} stale memories from vector store.`;
-    }
-    catch (err) {
-        logger.error(`[engram:settle] Step6 pruneStaleMemories failed: ${err}`);
-        return `Prune skipped: ${err}`;
-    }
-}
-/**
- * 执行每日沉淀流程（精简版）
- * v0.5: 每日 settle 只做：
- *   1. 结构化短期记忆
- *   2. 更新画像
- *   3. 生成画像摘要
- * 不再执行 extractAndVectorize（改为月度精选沉淀）和 archiveShortTerm（文件永久保留）
- *
- * @param opts 沉淀选项
- * @param targetDate 可选，指定要处理的日期 (YYYY-MM-DD)，默认当天
- */
-export async function runSettlement(opts, targetDate) {
-    const date = targetDate || getLocalDateString();
-    const results = [];
-    const startTime = Date.now();
-    logger.info(`[engram:settle] ====== Daily Settlement START for ${date} ======`);
-    logger.info(`[engram:settle] Config: shortTermDays=${opts.config.shortTermDays}, halfLifeDays=${opts.config.halfLifeDays}`);
-    logger.info(`[engram:settle] Has llmCall=${!!opts.llmCall}, has vectorStore=${!!opts.vectorStore}`);
-    // Step 1: 结构化短期记忆
-    const stepStart1 = Date.now();
-    results.push(await structurizeShortTerm(opts, date));
-    logger.info(`[engram:settle] Step1 (structurize) took ${Date.now() - stepStart1}ms: ${results[results.length - 1]}`);
-    // Step 2: 更新画像（基于新结构化的内容）
-    const stepStart2 = Date.now();
-    results.push(await updateProfile(opts, date));
-    logger.info(`[engram:settle] Step2 (profile update) took ${Date.now() - stepStart2}ms: ${results[results.length - 1]}`);
-    // Step 3: 生成画像摘要
-    const stepStart3 = Date.now();
-    results.push(await generateProfileSummary(opts));
-    logger.info(`[engram:settle] Step3 (profile summary) took ${Date.now() - stepStart3}ms: ${results[results.length - 1]}`);
-    // 注意：以下步骤已从每日 settle 中移除
-    // - extractAndVectorize → 改为月度精选沉淀 (monthlySettle)
-    // - archiveShortTerm → 短期记忆文件永久保留，不再归档删除
-    // - pruneStaleMemories → 改为月度遗忘 (monthlySettle)
-    logger.info(`[engram:settle] ====== Daily Settlement DONE (total ${Date.now() - startTime}ms) ======`);
-    return results;
-}
-/**
- * 月度沉淀 + 遗忘
- *
- * 两阶段执行：
- * 1. 精选沉淀：从 recall-hits.json 中筛选 hitCount ≥ minHitCount 的短期记忆，
- *    读取对应文件内容，调用 LLM 精简后写入向量库
- * 2. 月度遗忘：清理向量库中 effective_importance 低于阈值的长期记忆
- *
- * @param opts 沉淀选项（需要 llmCall + vectorStore + vectorPrune）
- * @param minHitCount 最低命中次数，默认 2（被召回 ≥2 次才沉淀）
- */
-export async function runMonthlySettle(opts, minHitCount = 2) {
-    const results = [];
-    const startTime = Date.now();
-    logger.info(`[engram:settle] ====== Monthly Settlement START ======`);
-    logger.info(`[engram:settle] minHitCount=${minHitCount}, has llmCall=${!!opts.llmCall}, has vectorStore=${!!opts.vectorStore}, has vectorPrune=${!!opts.vectorPrune}`);
-    // ---- Phase 1: 精选沉淀 ----
-    const phase1Start = Date.now();
-    const phase1Result = await monthlyExtractAndVectorize(opts, minHitCount);
-    results.push(phase1Result);
-    logger.info(`[engram:settle] Phase1 (extract) took ${Date.now() - phase1Start}ms: ${phase1Result}`);
-    // ---- Phase 2: 月度遗忘 ----
-    const phase2Start = Date.now();
-    const phase2Result = await pruneStaleMemories(opts);
-    results.push(phase2Result);
-    logger.info(`[engram:settle] Phase2 (prune) took ${Date.now() - phase2Start}ms: ${phase2Result}`);
-    logger.info(`[engram:settle] ====== Monthly Settlement DONE (total ${Date.now() - startTime}ms) ======`);
-    return results;
-}
-/**
- * 月度精选沉淀：只沉淀被召回命中 ≥ minHitCount 次的短期记忆
- *
- * 流程：
- * 1. 读取 recall-hits.json，筛选 hitCount >= minHitCount 且 settled === false
- * 2. 对筛选出的文件，读取内容并拼接
- * 3. 调用 LLM 提取值得长期保留的条目（复用 extractAndVectorize 的 prompt）
- * 4. 写入向量库
- * 5. 在 recall-hits.json 中标记 settled: true
- */
-async function monthlyExtractAndVectorize(opts, minHitCount) {
-    logger.info(`[engram:settle] monthlyExtractAndVectorize: minHitCount=${minHitCount}`);
-    if (!opts.llmCall || !opts.vectorStore) {
-        logger.info(`[engram:settle] monthlyExtractAndVectorize: missing llmCall=${!!opts.llmCall} vectorStore=${!!opts.vectorStore}`);
-        return "No LLM/vector configured.";
-    }
-    // 1. 读取 recall-hits.json
-    const recallHitsPath = join(opts.workspaceDir, "memory-engram", "recall-hits.json");
-    let hits;
-    try {
-        const raw = await fs.readFile(recallHitsPath, "utf-8");
-        const parsed = JSON.parse(raw);
-        // 兼容 v1 格式
-        hits = {};
-        for (const [key, value] of Object.entries(parsed)) {
-            if (typeof value === "number") {
-                hits[key] = { hitCount: 1, firstHit: value, lastHit: value, settled: false };
-            }
-            else if (typeof value === "object" && value !== null) {
-                hits[key] = value;
-            }
-        }
-    }
-    catch {
-        logger.info(`[engram:settle] monthlyExtractAndVectorize: no recall-hits.json, nothing to settle`);
-        return "No recall-hits.json found.";
-    }
-    // 2. 筛选：hitCount >= minHitCount 且 settled === false
-    const candidates = Object.entries(hits).filter(([_, entry]) => entry.hitCount >= minHitCount && !entry.settled);
-    logger.info(`[engram:settle] monthlyExtractAndVectorize: ${Object.keys(hits).length} total entries, ${candidates.length} meet criteria (hitCount≥${minHitCount}, not settled)`);
-    if (candidates.length === 0) {
-        return "No short-term memories meet monthly settle criteria.";
-    }
-    // 3. 读取对应的短期记忆文件内容
-    const shortTermDir = join(opts.workspaceDir, "memory-engram", "short-term");
-    const contentParts = [];
-    const processedFiles = [];
-    for (const [fileName, _hitEntry] of candidates) {
-        const filePath = join(shortTermDir, fileName);
-        try {
-            const content = await fs.readFile(filePath, "utf-8");
-            if (content.trim()) {
-                contentParts.push(`--- ${fileName} ---\n${content}`);
-                processedFiles.push(fileName);
-                logger.info(`[engram:settle] monthlyExtractAndVectorize: loaded ${fileName} (${content.length} chars, hitCount=${_hitEntry.hitCount})`);
-            }
-        }
-        catch {
-            logger.info(`[engram:settle] monthlyExtractAndVectorize: file ${fileName} not found, skipping`);
-        }
-    }
-    if (contentParts.length === 0) {
-        return "All candidate files missing or empty.";
-    }
-    const allContent = contentParts.join("\n\n");
-    logger.info(`[engram:settle] monthlyExtractAndVectorize: total content ${allContent.length} chars from ${processedFiles.length} files`);
-    // 4. 调用 LLM 提取值得长期保留的条目
-    const systemPrompt = `你是一个信息筛选助手。从以下短期记忆中筛选出值得长期保留的条目。
-
-这些短期记忆已经被多次召回命中，说明它们有持续价值。请提炼出核心信息。
-
-保留标准：
-- 包含用户偏好、决策、重要事件、具体信息（地点、金额、人名）
-- 包含用户行为模式和生活习惯（作息时间、运动习惯、饮食规律等）
-- 包含技术决策、项目配置、工作流程等持续有用的信息
-- 合并同一主题的多次记录为一条精炼表述
-- 去掉纯闲聊、临时信息、已过时的内容
-
-对每条保留的条目，输出 JSON 数组，每条包含：
-- text: 条目摘要（精炼、信息密度高）
-- category: 分类
-- importance: 重要度评分 0-1，标准如下：
-  - 0.9-1.0: 关键决策、重要事件（搬家、换工作、关系变化）
-  - 0.7-0.8: 明确偏好、项目里程碑、具体计划
-  - 0.4-0.6: 一般性讨论、技术问答
-  - 0.1-0.3: 日常闲聊、低价值内容
-
-只输出 JSON，不要其他内容。`;
-    logger.info(`[engram:settle] monthlyExtractAndVectorize: calling LLM for extraction (${allContent.length} chars)`);
-    const result = await opts.llmCall(allContent, systemPrompt);
-    logger.info(`[engram:settle] monthlyExtractAndVectorize: LLM returned ${result.length} chars`);
-    let items;
-    try {
-        const cleaned = cleanLLMJson(result);
-        items = JSON.parse(cleaned);
-        logger.info(`[engram:settle] monthlyExtractAndVectorize: parsed ${items.length} candidate items`);
-    }
-    catch (parseErr) {
-        logger.error(`[engram:settle] monthlyExtractAndVectorize: JSON parse failed: ${parseErr}, raw="${result.slice(0, 200)}"`);
-        return `Monthly settle skipped: JSON parse failed`;
-    }
-    // 5. 写入向量库
-    let stored = 0;
-    let duplicates = 0;
-    let skippedInvalid = 0;
-    for (const item of items) {
-        if (item.text && item.category) {
-            const imp = typeof item.importance === "number" ? item.importance : 0.5;
-            logger.info(`[engram:settle] monthlyExtractAndVectorize: vectorizing [${item.category}] importance=${imp.toFixed(2)} "${item.text.slice(0, 60)}"`);
-            const storeResult = await opts.vectorStore(item.text, item.category, imp);
-            if (storeResult === "created")
-                stored++;
-            else
-                duplicates++;
-        }
-        else {
-            skippedInvalid++;
-            logger.info(`[engram:settle] monthlyExtractAndVectorize: skipping invalid item: ${JSON.stringify(item).slice(0, 100)}`);
-        }
-    }
-    // 6. 标记 settled: true
-    for (const fileName of processedFiles) {
-        if (hits[fileName]) {
-            hits[fileName].settled = true;
-        }
-    }
-    await fs.writeFile(recallHitsPath, JSON.stringify(hits, null, 2), "utf-8");
-    logger.info(`[engram:settle] monthlyExtractAndVectorize: marked ${processedFiles.length} files as settled`);
-    const summary = `Monthly settle: ${stored} items stored (${duplicates} dup, ${skippedInvalid} invalid) from ${processedFiles.length} files.`;
-    logger.info(`[engram:settle] monthlyExtractAndVectorize: done. ${summary}`);
-    return summary;
-}
-//# sourceMappingURL=settle.js.map