@modusensus/dsh-mneme 0.4.5 → 0.4.6

package/test/quality-filter.test.js

@@ -0,0 +1,118 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import { createStore } from "../src/store.js";
+import { createService } from "../src/service.js";
+import {
+  evaluateMemoryQuality,
+  META_MEMORY_RE,
+  textSimilarity,
+  dedupRatio
+} from "../src/quality-filter.js";
+
+// Bug7: rule-based memory quality filter. Gated on
+// config.memoryQualityFilter.enabled === true — raw configs (`{}`) keep the
+// legacy behavior. Score bands:
+//   >= 60       stored normally
+//   30 .. < 60  quality_score persisted, injection ranked by importance × score/100
+//   < 30        archived + tagged low_quality (explicit search still recalls it)
+
+function setup(over = {}) {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: { memoryQualityFilter: { enabled: true }, ...over } });
+  return { store, service };
+}
+
+test("evaluateMemoryQuality: meta-memory scores below 60 (degraded, not archived)", () => {
+  const { score, tags } = evaluateMemoryQuality({
+    type: "history",
+    title: "对话总结",
+    content: "总结一下刚才的对话，需要记住以下要点"
+  });
+  assert.ok(score >= 30 && score < 60, `meta memory should be degraded (30..60), got ${score}`);
+  assert.ok(tags.includes("meta"), "tagged meta");
+});
+
+test("evaluateMemoryQuality: short content is archived (< 30)", () => {
+  const { score, tags } = evaluateMemoryQuality({ type: "preference", title: "语言", content: "短" });
+  assert.ok(score < 30, `short content should be archived, got ${score}`);
+  assert.ok(tags.includes("short_content"));
+  assert.ok(tags.includes("low_quality"));
+});
+
+test("evaluateMemoryQuality: near-duplicate of a recent memory is archived", () => {
+  const { score, tags } = evaluateMemoryQuality(
+    { type: "preference", title: "重复", content: "猫咪喜欢在阳台晒太阳并打盹" },
+    { recentContents: ["猫咪喜欢在阳台晒太阳并打盹"] }
+  );
+  assert.ok(score < 30, `duplicate should be archived, got ${score}`);
+  assert.ok(tags.includes("duplicate"));
+});
+
+test("saveWithDedupe: meta memory persists quality_score < 60 and is not archived", () => {
+  const { store, service } = setup();
+  const { memory } = service.saveWithDedupe({
+    type: "history",
+    title: "对话总结",
+    content: "总结一下刚才的对话，需要记住以下要点"
+  });
+  const got = store.getById(memory.id);
+  assert.ok(got.quality_score !== undefined && got.quality_score < 60,
+    `meta memory should store quality_score < 60, got ${got.quality_score}`);
+  assert.equal(got.archived, false, "degraded memory stays un-archived");
+});
+
+test("saveWithDedupe: short text is archived + tagged low_quality", () => {
+  const { store, service } = setup();
+  const { memory } = service.saveWithDedupe({
+    type: "preference",
+    title: "语言",
+    content: "短"
+  });
+  const got = store.getById(memory.id);
+  assert.equal(got.archived, true, "short text archived");
+  assert.ok(got.tags.includes("low_quality"), "tagged low_quality");
+  assert.ok(got.quality_score < 30, `score below archive threshold, got ${got.quality_score}`);
+});
+
+test("saveWithDedupe: near-duplicate content is archived", () => {
+  const { store, service } = setup();
+  service.saveWithDedupe({ type: "preference", title: "习惯一", content: "猫咪喜欢在阳台晒太阳并打盹" });
+  const { memory: dup } = service.saveWithDedupe({ type: "preference", title: "习惯二", content: "猫咪喜欢在阳台晒太阳并打盹" });
+  const got = store.getById(dup.id);
+  assert.equal(got.archived, true, "near-duplicate archived");
+  assert.ok(got.tags.includes("low_quality"));
+});
+
+test("saveWithDedupe: filter disabled (raw config) skips scoring entirely", () => {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: {} });
+  const { memory } = service.saveWithDedupe({ type: "preference", title: "语言", content: "短" });
+  const got = store.getById(memory.id);
+  assert.equal(got.archived, false, "nothing archived when filter off");
+  assert.equal(got.quality_score, undefined, "no quality_score when filter off");
+  assert.ok(!got.tags.includes("low_quality"));
+});
+
+test("injectCandidates: degraded memory ranks below a healthy preference", () => {
+  const { service } = setup();
+  // Healthy preference: score 100 → weight 1.0 → 3 * 1.0 = 3.0.
+  service.saveWithDedupe({ type: "preference", title: "健康偏好", content: "用户平时习惯用中文交流", importance: 3 });
+  // Degraded (meta-memory) preference: score ~55 → weight 0.55 → 5 * 0.55 = 2.75.
+  service.saveWithDedupe({ type: "preference", title: "元记忆偏好", content: "总结一下刚才的对话内容吧", importance: 5 });
+  const candidates = service.injectCandidates({ maxItems: 5, threshold: 3 });
+  const titles = candidates.map((c) => c.title);
+  assert.ok(titles.includes("健康偏好"), "healthy preference present");
+  assert.ok(titles.includes("元记忆偏好"), "degraded preference present");
+  assert.ok(titles.indexOf("健康偏好") < titles.indexOf("元记忆偏好"),
+    "degraded memory is demoted below the healthy one");
+});
+
+test("exported helpers behave (meta regex, similarity, dedup ratio)", () => {
+  assert.ok(META_MEMORY_RE.test("总结一下刚才的对话"));
+  assert.ok(META_MEMORY_RE.test("作为AI助手，我需要记住"));
+  assert.ok(!META_MEMORY_RE.test("用户喜欢喝咖啡"));
+  assert.equal(textSimilarity("猫咪喜欢晒太阳", "猫咪喜欢晒太阳"), 1);
+  assert.ok(textSimilarity("猫咪喜欢晒太阳", "完全不同的内容") < 0.3);
+  assert.ok(dedupRatio("哈哈哈哈哈哈") < 0.3, "repetitive filler has low dedup ratio");
+  assert.ok(dedupRatio("一句信息量足够的话") > 0.3);
+});

package/test/service.test.js

@@ -32,8 +32,15 @@ test("saveWithDedupe merges into existing on exact title match", () => {
   assert.equal(result.action, "merged");
   assert.equal(store.count(), 1);
   const all = store.all();
-  assert.equal(all[0].content, "新内容");
-  assert.equal(all[0].importance, 5);
+  // Bug5: same-title merge appends under a timestamped `---` separator instead
+  // of overwriting; the previous content is archived into content_history.
+  assert.ok(all[0].content.includes("旧内容"), "old content preserved in appended body");
+  assert.ok(all[0].content.includes("新内容"), "new content appended");
+  assert.ok(all[0].content.includes("---"), "timestamped separator present");
+  assert.equal(all[0].importance, 5, "importance takes the max of both");
+  assert.equal(all[0].content_history.length, 1, "old content archived to history");
+  assert.equal(all[0].content_history[0].content, "旧内容");
+  assert.equal(all[0].content_history[0].source, "auto_merge");
 });
 
 test("injectCandidates returns preferences + high-importance items within limit", () => {
@@ -196,3 +203,127 @@ test("service.compareAndUpdate rejects a stale version token (no lost update)",
   assert.equal(service.getById(m.id).content, "count=1", "increment not lost");
   store.close();
 });
+
+// --- Bug4: injectCandidates semantic-first recall (query) --------------------
+
+test("Bug4: injectCandidates with an empty query keeps the legacy rule-based selection", () => {
+  const { service } = setup();
+  service.saveWithDedupe({ type: "preference", title: "p0", content: "偏好" });
+  service.saveWithDedupe({ type: "preference", title: "p1", content: "偏好" });
+  service.saveWithDedupe({ type: "project", title: "low", content: "低", importance: 2 });
+  service.saveWithDedupe({ type: "project", title: "high", content: "高", importance: 5 });
+  service.saveWithDedupe({ type: "history", title: "h", content: "历史", importance: 5 });
+  service.saveWithDedupe({ type: "summary", title: "记忆库总览", content: "总览", importance: 5 });
+  const legacy = service.injectCandidates({ maxItems: 5, threshold: 4 });
+  const noQuery = service.injectCandidates({ maxItems: 5, threshold: 4, query: "" });
+  const blankQuery = service.injectCandidates({ maxItems: 5, threshold: 4, query: "   " });
+  assert.deepEqual(noQuery.map((c) => c.id), legacy.map((c) => c.id), "empty query behaves identically to the old logic");
+  assert.deepEqual(blankQuery.map((c) => c.id), legacy.map((c) => c.id), "whitespace-only query degrades too");
+  assert.equal(legacy[0].type, "summary", "summary still leads");
+  assert.ok(!legacy.some((c) => c.type === "history"), "history still excluded");
+});
+
+test("Bug4: injectCandidates ignores the query when hybridInject is disabled", () => {
+  const { store } = setup();
+  const svc = createService({ store, mirror: null, config: { hybridInject: false } });
+  svc.saveWithDedupe({ type: "preference", title: "p0", content: "偏好" });
+  svc.saveWithDedupe({ type: "summary", title: "记忆库总览", content: "总览", importance: 5 });
+  const noQuery = svc.injectCandidates({ maxItems: 5, threshold: 3 });
+  const withQuery = svc.injectCandidates({ maxItems: 5, threshold: 3, query: "中文" });
+  assert.deepEqual(withQuery.map((c) => c.id), noQuery.map((c) => c.id), "hybridInject off keeps rule-based ordering");
+});
+
+// --- Bug5: content_history FIFO + human-edited direct overwrite --------------
+
+test("Bug5: content_history grows FIFO across repeated same-title merges (capped at 20)", () => {
+  const { store, service } = setup();
+  service.saveWithDedupe({ type: "project", title: "流水", content: "v0" });
+  for (let i = 1; i <= 25; i++) {
+    service.saveWithDedupe({ type: "project", title: "流水", content: `v${i}` });
+  }
+  const row = store.all()[0];
+  assert.ok(Array.isArray(row.content_history), "content_history is an array");
+  assert.equal(row.content_history.length, 20, "FIFO cap of 20 respected");
+  assert.ok(row.content_history[0].content.includes("v24"), "newest archived body kept at index 0");
+  assert.ok(row.content_history[19].content.includes("v5"), "oldest kept entry");
+  assert.ok(!row.content_history[19].content.includes("v6"), "entries older than the cap dropped");
+  assert.ok(row.content.includes("v25"), "latest merge still appended to the live body");
+});
+
+test("Bug5: _humanEdited overwrite replaces content directly and archives the old version", () => {
+  const { store, service } = setup();
+  service.saveWithDedupe({ type: "preference", title: "语言", content: "机器内容" });
+  const result = service.saveWithDedupe({ type: "preference", title: "语言", content: "人工修正", _humanEdited: true });
+  const row = store.getById(result.memory.id);
+  assert.equal(row.content, "人工修正", "direct overwrite, no appended --- block");
+  assert.ok(!row.content.includes("机器内容"), "old content not in the live body");
+  assert.equal(row.content_history.length, 1);
+  assert.equal(row.content_history[0].content, "机器内容");
+  assert.equal(row.content_history[0].source, "human_override");
+});
+
+// --- Bug7: memory quality filter ---------------------------------------------
+
+function qualitySetup(config = {}) {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: { memoryQualityFilter: { enabled: true, ...config } } });
+  return { store, service };
+}
+
+test("Bug7: meta-memory is scored below the degrade threshold but still stored", () => {
+  const { store, service } = qualitySetup();
+  const result = service.saveWithDedupe({ type: "project", title: "记忆系统规则", content: "我需要记住用户喜欢中文阅读长篇小说和散文" });
+  assert.equal(result.action, "created");
+  assert.ok(result.memory.quality_score < 60, `meta-memory degraded, got ${result.memory.quality_score}`);
+  assert.ok(result.memory.quality_score >= 30, "still above the archive threshold");
+  assert.equal(result.memory.archived, false, "not archived");
+  assert.equal(store.count(), 1, "still stored and searchable");
+});
+
+test("Bug7: short content is archived and tagged low_quality", () => {
+  const { store, service } = qualitySetup();
+  const result = service.saveWithDedupe({ type: "project", title: "备忘", content: "x" });
+  const row = store.getById(result.memory.id);
+  assert.equal(row.archived, true, "short content archived");
+  assert.ok(row.tags.includes("low_quality"), "low_quality tag added");
+  assert.ok(row.quality_score < 30, `score below archive threshold: ${row.quality_score}`);
+});
+
+test("Bug7: near-duplicate content is archived and tagged low_quality", () => {
+  const { store, service } = qualitySetup();
+  const base = "今天下午去操场跑步三圈然后回来洗澡";
+  service.saveWithDedupe({ type: "history", title: "记录A", content: base });
+  const result = service.saveWithDedupe({ type: "history", title: "记录B", content: `${base}${base}${base}` });
+  const row = store.getById(result.memory.id);
+  assert.equal(row.archived, true, "near-duplicate archived");
+  assert.ok(row.tags.includes("low_quality"), "low_quality tag added");
+  assert.ok(row.quality_score < 30, `score below archive threshold: ${row.quality_score}`);
+});
+
+test("Bug7: repetitive (dedup<0.3) content is scored into the degraded band", () => {
+  const { store, service } = qualitySetup();
+  const result = service.saveWithDedupe({ type: "history", title: "唠叨", content: "好好好好好好好好好好好好好好好好好好" });
+  const row = store.getById(result.memory.id);
+  assert.ok(row.quality_score >= 30 && row.quality_score < 60, `repetitive-only degraded, got ${row.quality_score}`);
+  assert.equal(row.archived, false, "not archived (only a single non-meta signal)");
+});
+
+test("Bug7: memoryQualityFilter.enabled=false skips scoring entirely", () => {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: { memoryQualityFilter: { enabled: false } } });
+  const result = service.saveWithDedupe({ type: "project", title: "记忆系统规则", content: "我需要记住用户喜欢中文阅读长篇小说和散文" });
+  const row = store.getById(result.memory.id);
+  assert.ok(row.quality_score == null, "no quality_score written");
+  assert.equal(row.archived, false, "not archived");
+  assert.ok(!(row.tags ?? []).includes("low_quality"), "no low_quality tag");
+});
+
+test("Bug7: injection ranking re-weights by quality_score (degraded memories rank lower)", () => {
+  const { store } = setup();
+  const svc = createService({ store, mirror: null, config: { memoryQualityFilter: { enabled: true } } });
+  svc.saveWithDedupe({ type: "preference", title: "高质量", content: "用户喜欢读科幻小说和散文" });
+  svc.saveWithDedupe({ type: "preference", title: "元记忆", content: "我需要记住用户偏好的完整清单防止忘记" });
+  const candidates = svc.injectCandidates({ maxItems: 5, threshold: 3 });
+  assert.equal(candidates[0].title, "高质量", "100-quality preference leads the degraded one");
+  assert.equal(candidates[1].title, "元记忆");
+});

package/test/vector-index.test.js

@@ -21,9 +21,9 @@ function addMemory(store, { title, content, type = "preference", vector }) {
   return memory;
 }
 
-// Embedder mock for rebuildIndex. NOTE: the rebuild guard checks
-// `typeof embedder.embed === "function"` while the inner loop calls
-// `embedder.embedSingle` — so a rebuild-ready mock must provide both.
+// Embedder mock for rebuildIndex. The rebuild loop prefers embedSingle and
+// falls back to `embed` for embed-only OpenAI-compatible clients, so a mock
+// that provides embedSingle exercises the primary path.
 function rebuildEmbedder({ vector = [1, 0, 0], modelHash = "mock#rebuild", dimension = 3 } = {}) {
   return {
     embed: async (texts) => texts.map(() => vector),
@@ -137,15 +137,31 @@ test("rebuildIndex handles a throwing embedder without failing the batch", async
   assert.equal(vectorIndex.getStats().embeddedCount, 1);
 });
 
-test("rebuildIndex guard: embedder without embedSingle() is ignored", async () => {
+test("rebuildIndex guard: embedder with no embed method at all is ignored", async () => {
   const { store, vectorIndex } = setup();
   addMemory(store, { title: "缺向量", content: "A" });
-  const bare = { embed: async () => [1, 0, 0], modelHash: "mock#bare", dimension: 3 };
+  const bare = { modelHash: "mock#bare", dimension: 3 };
   const result = await vectorIndex.rebuildIndex(bare, { limit: 100 });
-  assert.deepEqual(result, { indexed: 0, skipped: 0 }, "guard returns early when embedSingle() is missing");
+  assert.deepEqual(result, { indexed: 0, skipped: 0 }, "guard returns early when neither embed nor embedSingle is available");
   assert.equal(vectorIndex.modelHash(), undefined, "no model marked");
 });
 
+test("rebuildIndex supports an embed-only OpenAI-compatible embedder (issue #10)", async () => {
+  const { store, vectorIndex } = setup();
+  addMemory(store, { title: "缺向量", content: "A" });
+  const embedOnly = {
+    embed: async (text) => [1, 0, 0], // OpenAI-compatible single-text embed
+    modelHash: "text-embedding#abc",
+    dimension: 3
+  };
+  const result = await vectorIndex.rebuildIndex(embedOnly, { limit: 100 });
+  assert.equal(result.indexed, 1, "embed-only embedder indexes the missing row");
+  assert.equal(result.skipped, 0);
+  assert.equal(vectorIndex.modelHash(), "text-embedding#abc", "model fingerprint recorded");
+  assert.equal(vectorIndex.dimension(), 3, "dimension recorded");
+  assert.equal(vectorIndex.getStats().embeddedCount, 1, "embedding persisted");
+});
+
 test("rebuildIndex works with embedSingle-only embedder", async () => {
   const { store, vectorIndex } = setup();
   addMemory(store, { title: "缺向量", content: "A" });