@modusensus/dsh-mneme 0.4.5 → 0.4.6

package/lib/store.js

@@ -12,6 +12,7 @@ CREATE TABLE IF NOT EXISTS memories (
   forgotten   INTEGER NOT NULL DEFAULT 0,
   archived    INTEGER NOT NULL DEFAULT 0,
   source      TEXT,
+  content_history TEXT,
   embedding   TEXT,
   epistemic_status TEXT NOT NULL DEFAULT 'subjective',
   last_accessed_at  TEXT,
@@ -146,6 +147,31 @@ CREATE TABLE IF NOT EXISTS conflict_pending (
 );
 CREATE INDEX IF NOT EXISTS idx_conflict_pending_unresolved ON conflict_pending(resolved_at);
 
+-- llm_audit_logs: every background LLM call (autoDream consolidation + summary,
+-- autoSummarize compression) is recorded here — tokens in/out, duration, status
+-- and the trigger that caused it (Bug8). Failures are captured as status='error'
+-- and never block the calling feature. retentionDays is enforced by a boot-time
+-- purge (deleteOldLlmAudits). Bookkeeping like the other audit tables: it never
+-- triggers write hooks.
+CREATE TABLE IF NOT EXISTS llm_audit_logs (
+  id                INTEGER PRIMARY KEY AUTOINCREMENT,
+  timestamp         TEXT NOT NULL,
+  trigger_source    TEXT NOT NULL,          -- autoDream | autoSummarize | manual ...
+  operation_type    TEXT NOT NULL,          -- dream_consolidate | dream_summarize | summarize_compress ...
+  model_id          TEXT NOT NULL,
+  input_tokens      INTEGER NOT NULL DEFAULT 0,
+  output_tokens     INTEGER NOT NULL DEFAULT 0,
+  total_tokens      INTEGER NOT NULL DEFAULT 0,
+  cost_usd          REAL NOT NULL DEFAULT 0,
+  duration_ms       INTEGER NOT NULL DEFAULT 0,
+  status            TEXT NOT NULL,          -- success | error | skipped
+  error_message     TEXT,
+  related_memory_ids TEXT,                  -- JSON: ids the call operated on
+  metadata          TEXT                    -- JSON: free-form extras
+);
+CREATE INDEX IF NOT EXISTS idx_llm_audit_timestamp ON llm_audit_logs(timestamp);
+CREATE INDEX IF NOT EXISTS idx_llm_audit_source ON llm_audit_logs(trigger_source);
+
 -- entity gene (v0.3.0): named entities mentioned across memories, with
 -- time-boxed attributes (valid_from → valid_until) and typed relations.
 -- Attributes follow the snapshot style: saveAttr invalidates the previous
@@ -295,6 +321,8 @@ function toRow(row) {
     forgotten: row.forgotten === 1,
     archived: row.archived === 1,
     source: row.source ?? undefined,
+    content_history: parseJsonArray(row.content_history),
+    quality_score: row.quality_score !== null && row.quality_score !== undefined ? Number(row.quality_score) : undefined,
     epistemic_status: row.epistemic_status ?? "subjective",
     created_at: row.created_at,
     updated_at: row.updated_at,
@@ -439,6 +467,34 @@ function toRelation(row) {
   };
 }
 
+function toLlmAudit(row) {
+  if (!row) return undefined;
+  let metadata;
+  if (row.metadata != null) {
+    try {
+      metadata = JSON.parse(row.metadata);
+    } catch {
+      metadata = row.metadata;
+    }
+  }
+  return {
+    id: row.id,
+    timestamp: row.timestamp,
+    trigger_source: row.trigger_source,
+    operation_type: row.operation_type,
+    model_id: row.model_id,
+    input_tokens: row.input_tokens,
+    output_tokens: row.output_tokens,
+    total_tokens: row.total_tokens,
+    cost_usd: row.cost_usd,
+    duration_ms: row.duration_ms,
+    status: row.status,
+    error_message: row.error_message ?? undefined,
+    related_memory_ids: parseJsonArray(row.related_memory_ids),
+    metadata
+  };
+}
+
 function toMirrorState(row) {
   if (!row) {
     return {
@@ -509,6 +565,12 @@ export function createStore(path) {
   if (!columns.includes("epistemic_status")) {
     db.exec("ALTER TABLE memories ADD COLUMN epistemic_status TEXT NOT NULL DEFAULT 'subjective'");
   }
+  if (!columns.includes("content_history")) {
+    db.exec("ALTER TABLE memories ADD COLUMN content_history TEXT");
+  }
+  if (!columns.includes("quality_score")) {
+    db.exec("ALTER TABLE memories ADD COLUMN quality_score REAL");
+  }
 
   // Legacy dream_runs without policy_epoch → backfill with the default epoch.
   const dreamCols = db.prepare("PRAGMA table_info(dream_runs)").all().map((c) => c.name);
@@ -607,9 +669,24 @@ export function createStore(path) {
       : inferEpistemicStatus(memory);
     runAtomically(() => {
       db.prepare(
-        `INSERT INTO memories (id, type, title, content, tags, importance, forgotten, source, embedding, epistemic_status, created_at, updated_at)
-         VALUES (?, ?, ?, ?, ?, ?, 0, ?, ?, ?, ?, ?)`
-      ).run(id, type, memory.title, memory.content, tags, importance, memory.source ?? null, embedding, epistemicStatus, now, now);
+        `INSERT INTO memories (id, type, title, content, tags, importance, forgotten, archived, source, content_history, quality_score, embedding, epistemic_status, created_at, updated_at)
+         VALUES (?, ?, ?, ?, ?, ?, 0, ?, ?, ?, ?, ?, ?, ?, ?)`
+      ).run(
+        id,
+        type,
+        memory.title,
+        memory.content,
+        tags,
+        importance,
+        memory.archived ? 1 : 0,
+        memory.source ?? null,
+        JSON.stringify(memory.content_history ?? []),
+        Number.isFinite(memory.quality_score) ? memory.quality_score : null,
+        embedding,
+        epistemicStatus,
+        now,
+        now
+      );
       // desired generation bumped in the same transaction as the write: once
       // this commits, generation > applied_generation, so a crash right after
       // (before syncMirror) is caught by recoverMirror on restart (peer
@@ -632,9 +709,15 @@ export function createStore(path) {
       ? (Array.isArray(patch.embedding) && patch.embedding.length ? JSON.stringify(patch.embedding) : null)
       : existing.embedding ?? null;
     const epistemicStatus = resolveEpistemicStatus(existing, patch);
+    const contentHistory = Array.isArray(patch.content_history)
+      ? JSON.stringify(patch.content_history)
+      : (Array.isArray(existing.content_history) ? JSON.stringify(existing.content_history) : null);
+    const qualityScore = patch.quality_score !== undefined && Number.isFinite(patch.quality_score)
+      ? patch.quality_score
+      : (existing.quality_score ?? null);
     runAtomically(() => {
       db.prepare(
-        `UPDATE memories SET type=?, title=?, content=?, tags=?, importance=?, source=?, embedding=?, epistemic_status=?, updated_at=? WHERE id=?`
+        `UPDATE memories SET type=?, title=?, content=?, tags=?, importance=?, source=?, content_history=?, quality_score=?, embedding=?, epistemic_status=?, updated_at=? WHERE id=?`
       ).run(
         type,
         patch.title ?? existing.title,
@@ -642,6 +725,8 @@ export function createStore(path) {
         JSON.stringify(patch.tags ?? existing.tags),
         Number.isInteger(patch.importance) ? patch.importance : existing.importance,
         patch.source !== undefined ? patch.source : (existing.source ?? null),
+        contentHistory,
+        qualityScore,
         embedding,
         epistemicStatus,
         now,
@@ -684,6 +769,12 @@ export function createStore(path) {
       ? (Array.isArray(patch.embedding) && patch.embedding.length ? JSON.stringify(patch.embedding) : null)
       : existing.embedding ?? null;
     const epistemicStatus = resolveEpistemicStatus(existing, patch);
+    const contentHistory = Array.isArray(patch.content_history)
+      ? JSON.stringify(patch.content_history)
+      : (Array.isArray(existing.content_history) ? JSON.stringify(existing.content_history) : null);
+    const qualityScore = patch.quality_score !== undefined && Number.isFinite(patch.quality_score)
+      ? patch.quality_score
+      : (existing.quality_score ?? null);
     // The CAS UPDATE and the desired-generation bump must commit together (audit
     // peer A): if the UPDATE autocommits first and the process dies before the
     // increment, the store is mutated while generation == applied_generation and
@@ -693,7 +784,7 @@ export function createStore(path) {
     let applied = false;
     runAtomically(() => {
       const result = db.prepare(
-        `UPDATE memories SET type=?, title=?, content=?, tags=?, importance=?, source=?, embedding=?, epistemic_status=?, updated_at=?
+        `UPDATE memories SET type=?, title=?, content=?, tags=?, importance=?, source=?, content_history=?, quality_score=?, embedding=?, epistemic_status=?, updated_at=?
          WHERE id=? AND updated_at=?`
       ).run(
         type,
@@ -702,6 +793,8 @@ export function createStore(path) {
         JSON.stringify(patch.tags ?? existing.tags),
         Number.isInteger(patch.importance) ? patch.importance : existing.importance,
         patch.source !== undefined ? patch.source : (existing.source ?? null),
+        contentHistory,
+        qualityScore,
         embedding,
         epistemicStatus,
         now,
@@ -1152,6 +1245,115 @@ export function createStore(path) {
     return rows.map(toRecallEval);
   }
 
+  // --- llm audit trail (Bug8) ---------------------------------------------
+
+  /**
+   * Persist one LLM audit row (a background call's token/time/status receipt).
+   * Bookkeeping like the other audit tables: it never triggers write hooks, so
+   * recording a call can never loop back into the scheduler that made it. The
+   * call itself is wrapped so a failure is captured (status='error') instead of
+   * blocking the feature — only a throwing saveLlmAudit is swallowed, never the
+   * LLM call.
+   */
+  function saveLlmAudit(entry) {
+    const now = nowIso();
+    const inTokens = Number.isFinite(entry.input_tokens) ? entry.input_tokens : 0;
+    const outTokens = Number.isFinite(entry.output_tokens) ? entry.output_tokens : 0;
+    db.prepare(
+      `INSERT INTO llm_audit_logs (timestamp, trigger_source, operation_type, model_id,
+        input_tokens, output_tokens, total_tokens, cost_usd, duration_ms, status,
+        error_message, related_memory_ids, metadata)
+       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`
+    ).run(
+      entry.timestamp ?? now,
+      entry.trigger_source,
+      entry.operation_type,
+      entry.model_id,
+      inTokens,
+      outTokens,
+      Number.isFinite(entry.total_tokens) ? entry.total_tokens : inTokens + outTokens,
+      Number.isFinite(entry.cost_usd) ? entry.cost_usd : 0,
+      Number.isFinite(entry.duration_ms) ? entry.duration_ms : 0,
+      entry.status ?? "success",
+      entry.error_message ?? null,
+      JSON.stringify(entry.related_memory_ids ?? []),
+      entry.metadata !== undefined
+        ? (typeof entry.metadata === "string" ? entry.metadata : JSON.stringify(entry.metadata))
+        : null
+    );
+    return toLlmAudit(db.prepare("SELECT * FROM llm_audit_logs ORDER BY id DESC LIMIT 1").get());
+  }
+
+  function listLlmAudits({ limit = 50, offset = 0, source } = {}) {
+    const { limit: lim, offset: off } = sanitizePage(limit, offset, 50);
+    const clauses = [];
+    const params = [];
+    if (source) {
+      clauses.push("trigger_source = ?");
+      params.push(source);
+    }
+    const where = clauses.length ? `WHERE ${clauses.join(" AND ")}` : "";
+    const rows = db.prepare(
+      `SELECT * FROM llm_audit_logs ${where} ORDER BY timestamp DESC, id DESC LIMIT ? OFFSET ?`
+    ).all(...params, lim, off);
+    return rows.map(toLlmAudit);
+  }
+
+  function countLlmAudits({ source } = {}) {
+    const clauses = [];
+    const params = [];
+    if (source) {
+      clauses.push("trigger_source = ?");
+      params.push(source);
+    }
+    const where = clauses.length ? `WHERE ${clauses.join(" AND ")}` : "";
+    return db.prepare(`SELECT count(*) AS c FROM llm_audit_logs ${where}`).get(...params).c;
+  }
+
+  /**
+   * Aggregate LLM spend over the last `days`: total calls/tokens/duration/cost,
+   * broken down by trigger_source and by status. Used by the API's
+   * /llm-audit/stats endpoint so the Web panel can show where budget goes.
+   */
+  function getLlmAuditStats({ days = 7 } = {}) {
+    const since = new Date(Date.now() - days * 86400000).toISOString();
+    const total = db.prepare(
+      `SELECT count(*) AS c,
+              COALESCE(SUM(input_tokens), 0) AS i,
+              COALESCE(SUM(output_tokens), 0) AS o,
+              COALESCE(SUM(total_tokens), 0) AS t,
+              COALESCE(SUM(duration_ms), 0) AS d,
+              COALESCE(SUM(cost_usd), 0) AS cst
+       FROM llm_audit_logs WHERE timestamp >= ?`
+    ).get(since);
+    const bySource = db.prepare(
+      `SELECT trigger_source AS source, count(*) AS c,
+              COALESCE(SUM(total_tokens), 0) AS total_tokens
+       FROM llm_audit_logs WHERE timestamp >= ?
+       GROUP BY trigger_source ORDER BY total_tokens DESC`
+    ).all(since);
+    const byStatus = db.prepare(
+      "SELECT status, count(*) AS c FROM llm_audit_logs WHERE timestamp >= ? GROUP BY status"
+    ).all(since);
+    return {
+      days,
+      since,
+      total_calls: total.c,
+      input_tokens: total.i,
+      output_tokens: total.o,
+      total_tokens: total.t,
+      total_duration_ms: total.d,
+      total_cost_usd: Number(total.cst),
+      by_source: bySource,
+      by_status: byStatus
+    };
+  }
+
+  /** Delete audit rows older than `before` (ISO string). Returns count removed. */
+  function deleteOldLlmAudits(before) {
+    return db.prepare("DELETE FROM llm_audit_logs WHERE timestamp < ?").run(before).changes;
+  }
+
   // --- failure memories ----------------------------------------------------
 
   /**
@@ -1678,6 +1880,11 @@ export function createStore(path) {
     saveRecallEval,
     getRecallEval,
     listRecallEvals,
+    saveLlmAudit,
+    listLlmAudits,
+    countLlmAudits,
+    getLlmAuditStats,
+    deleteOldLlmAudits,
     saveFailure,
     listFailures,
     getFailureStats,

package/lib/summarize.js

@@ -98,6 +98,12 @@ export function createSummarizer(ctx, service, config) {
     if (disposed || inFlight.has(session.id)) return;
     const controller = new AbortController();
     inFlight.set(session.id, controller);
+    // Bug8: audit state for the compression call. null = no audit for this run
+    // (disabled, or no LLM call was actually made). The audit row is written in
+    // the finally below — once, regardless of which exit path the call took —
+    // so a failed/aborted stream still leaves a status='error' trail without
+    // ever blocking the summarization itself.
+    let audit = null;
     try {
       const header = session.requestHeader?.()?.config;
       // Config override takes priority, then session header, then nothing.
@@ -110,6 +116,18 @@ export function createSummarizer(ctx, service, config) {
       const messages = collectMessages(session);
       if (!messages.length) return;
 
+      if (config?.llmAudit?.enabled !== false && typeof service.saveLlmAudit === "function") {
+        audit = {
+          route,
+          timestamp: new Date().toISOString(),
+          startedAt: Date.now(),
+          inputTokens: 0,
+          outputTokens: 0,
+          status: "success",
+          errorMessage: null
+        };
+      }
+
       const assembler = new BlockAssembler();
       let text = "";
       const options = {
@@ -122,15 +140,35 @@ export function createSummarizer(ctx, service, config) {
         ],
         signal: controller.signal
       };
-      for await (const chunk of ctx.llm.stream(options)) {
-        if (STREAM_CHUNK_TYPES.has(chunk.type)) assembler.push(toProtocolChunk(chunk));
-        if (chunk.type === "text-delta") {
-          text += chunk.text ?? chunk.delta ?? "";
+      try {
+        for await (const chunk of ctx.llm.stream(options)) {
+          if (STREAM_CHUNK_TYPES.has(chunk.type)) assembler.push(toProtocolChunk(chunk));
+          if (chunk.type === "text-delta") {
+            text += chunk.text ?? chunk.delta ?? "";
+          }
+          if (chunk.type === "usage" && audit) {
+            const i = chunk.input_tokens ?? chunk.inputTokens ?? chunk.prompt_tokens ?? chunk.promptTokens;
+            const o = chunk.output_tokens ?? chunk.outputTokens ?? chunk.completion_tokens ?? chunk.completionTokens;
+            if (Number.isFinite(i)) audit.inputTokens = i;
+            if (Number.isFinite(o)) audit.outputTokens = o;
+          }
+          if (chunk.type === "finish") {
+            const reasonKind = chunk.reason?.kind ?? chunk.kind;
+            if (reasonKind === "error" || reasonKind === "aborted") {
+              if (audit) {
+                audit.status = "error";
+                audit.errorMessage = `llm stream ${reasonKind}`;
+              }
+              return;
+            }
+          }
         }
-        if (chunk.type === "finish") {
-          const reasonKind = chunk.reason?.kind ?? chunk.kind;
-          if (reasonKind === "error" || reasonKind === "aborted") return;
+      } catch (error) {
+        if (audit) {
+          audit.status = "error";
+          audit.errorMessage = String(error?.message ?? error);
         }
+        throw error; // caller's catch handles the failure; audit already staged
       }
       // Direct delta accumulation is the primary extraction path (it works
       // for real protocol chunks {index,text} and looser {delta} shapes
@@ -147,6 +185,26 @@ export function createSummarizer(ctx, service, config) {
         service.saveWithDedupe({ ...entry, source: `session:${session.id}` });
       }
     } finally {
+      if (audit) {
+        try {
+          service.saveLlmAudit({
+            timestamp: audit.timestamp,
+            trigger_source: "autoSummarize",
+            operation_type: "summarize_compress",
+            model_id: `${audit.route.provider}:${audit.route.model}`,
+            input_tokens: audit.inputTokens,
+            output_tokens: audit.outputTokens,
+            total_tokens: audit.inputTokens + audit.outputTokens,
+            cost_usd: 0,
+            duration_ms: Date.now() - audit.startedAt,
+            status: audit.status,
+            error_message: audit.errorMessage,
+            related_memory_ids: []
+          });
+        } catch (auditError) {
+          ctx.logger?.warn?.(`dsh-mneme: llm audit write failed: ${String(auditError)}`);
+        }
+      }
       inFlight.delete(session.id);
     }
   }

package/lib/vector-index.js

@@ -73,13 +73,23 @@ export function createVectorIndex({ store, logger }) {
 
     /** Re-embed every row missing an embedding. Returns indexed count. */
     async rebuildIndex(embedder, { limit = 1000 } = {}) {
-      if (!embedder || typeof embedder.embedSingle !== "function") return { indexed: 0, skipped: 0 };
+      // The loop needs a single-text embedder. Native embedSingle is preferred;
+      // embed-only OpenAI-compatible clients (issue #10) are accepted too via
+      // their `embed` single-text interface, so /vector-reindex no longer
+      // silently returns 0 for them. An embedder exposing neither is ignored.
+      let embedOne = null;
+      if (embedder && typeof embedder.embedSingle === "function") {
+        embedOne = (text) => embedder.embedSingle(text);
+      } else if (embedder && typeof embedder.embed === "function") {
+        embedOne = (text) => Promise.resolve(embedder.embed(text));
+      }
+      if (!embedOne) return { indexed: 0, skipped: 0 };
       const rows = store.needsEmbedding(limit);
       let indexed = 0;
       for (const row of rows) {
         try {
           const text = [row.title, row.content].filter(Boolean).join("\n");
-          const vector = await embedder.embedSingle(text);
+          const vector = await embedOne(text);
           if (vector && vector.length) {
             store.setEmbedding(row.id, vector);
             indexed++;

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@modusensus/dsh-mneme",
   "description": "Cross-session memory plugin for DeepSeek Harness with autoDream consolidation: SQLite store, Markdown mirrors, 7 model tools, automatic injection, session summarization, user profile/rules, custom slash commands, vector (semantic) search, and a Web GUI panel",
-  "version": "0.4.5",
+  "version": "0.4.6",
   "license": "MIT",
   "type": "module",
   "main": "lib/index.js",

package/src/api.js

@@ -237,7 +237,9 @@ export function createApi(ctx, service, settings, commands, embedder, semantic =
         const task = viaIndex
           ? semantic.vectorIndex.rebuildIndex(embedder, { limit })
           : embedder.reindexMissing ? embedder.reindexMissing(limit) : Promise.resolve({ indexed: 0, skipped: 0, error: "vector-unavailable" });
-        task.then((result) => {
+        // Return the chain so awaiting callers (tests/health checks) observe the
+        // finished response rather than racing the async backfill.
+        return task.then((result) => {
           sendJson(res, 200, result);
         }).catch(() => {
           sendJson(res, 200, { indexed: 0, skipped: 0, error: "vector-failed" });
@@ -268,6 +270,42 @@ export function createApi(ctx, service, settings, commands, embedder, semantic =
     }
   });
 
+  // --- LLM audit trail (Bug8): paginated read + aggregate stats ---
+  // Read-only endpoints, so like list/search/semantic they stay open even when
+  // apiToken is set. The stats aggregate budget by source over the last N days.
+  register({
+    kind: "exact",
+    path: "/api/dsh-mneme/semantic/llm-audit",
+    handler(req, res) {
+      try {
+        const url = new URL(req.url, "http://localhost");
+        const page = Math.max(1, Number(url.searchParams.get("page") ?? 1) || 1);
+        const pageSize = Math.min(200, Math.max(1, Number(url.searchParams.get("pageSize") ?? 50) || 50));
+        const source = url.searchParams.get("source") ?? undefined;
+        const items = service.listLlmAudits?.({ limit: pageSize, offset: (page - 1) * pageSize, source }) ?? [];
+        const total = service.countLlmAudits?.({ source }) ?? items.length;
+        sendJson(res, 200, { items, total, page, pageSize });
+      } catch {
+        sendJson(res, 500, { error: "internal" });
+      }
+    }
+  });
+
+  register({
+    kind: "exact",
+    path: "/api/dsh-mneme/semantic/llm-audit/stats",
+    handler(req, res) {
+      try {
+        const url = new URL(req.url, "http://localhost");
+        const days = Math.max(1, Math.min(365, Number(url.searchParams.get("days") ?? 7) || 7));
+        const stats = service.getLlmAuditStats?.({ days }) ?? null;
+        sendJson(res, 200, stats ?? { error: "unavailable" });
+      } catch {
+        sendJson(res, 500, { error: "internal" });
+      }
+    }
+  });
+
   // --- health: mirror sync state (F-NEW-03 / v0.3.6) ---
   // Auth-gated; only returns a sanitized error code (never raw last_error which
   // may leak paths/token-like strings/internal hosts). On state read failure it
@@ -354,7 +392,7 @@ export function createApi(ctx, service, settings, commands, embedder, semantic =
   });
 
   return {
-    routes: 9,
+    routes: 11,
     dispose: () => {
       for (const dispose of disposers) dispose();
     }

package/src/config.js

@@ -82,6 +82,15 @@ export const Config = z.object({
   vectorSearchThreshold: z.number().min(0).max(1).default(0.65),
   hybridSearchVectorWeight: z.number().min(0).max(1).default(0.6),
   hybridSearchKeywordWeight: z.number().min(0).max(1).default(0.4),
+  // Lazy auto-backfill of missing embeddings on boot (Bug2): when the vector
+  // API is configured and rows still lack an embedding, the index is rebuilt
+  // in the background after a short delay, rate-limited in batches. On by
+  // default; set false to keep the backfill manual only.
+  autoReindexOnBoot: z.boolean().default(true),
+  // Semantic-first injection (Bug4): when enabled, injectCandidates with a
+  // non-empty query recalls via the vector index first and falls back to the
+  // rule-based pick to fill/dedupe. Empty query / no vector → legacy behavior.
+  hybridInject: z.boolean().default(true),
 
   // --- semantic: rerank layer (v0.2) --------------------------------------
   // Opt-in by default (item ⑥): the local cross-encoder pulls in onnxruntime
@@ -177,6 +186,32 @@ export const Config = z.object({
   // never used to influence behavior).
   trustEpistemicWeighting: z.boolean().default(false),
 
+  // --- memory quality filter (Bug7) ------------------------------------------
+  // Heuristic gate on what deserves the injection/recall surface. When enabled,
+  // saveWithDedupe scores each new memory after dedupe and before write:
+  //   score >= degradeThreshold (60) → stored normally
+  //   archiveThreshold (30) <= score < 60 → quality_score persisted and the
+  //       injection sort re-ranks by importance * quality_score/100 (degraded)
+  //   score < 30 → archived + tagged low_quality (still explicitly searchable)
+  // Meta-memory markers, near-duplicates and repetitive filler lose points.
+  memoryQualityFilter: z.object({
+    enabled: z.boolean().default(true),
+    archiveThreshold: z.natural().min(1).max(100).default(30),
+    degradeThreshold: z.natural().min(1).max(100).default(60),
+    minContentLength: z.natural().min(1).max(1000).default(10)
+  }).default({}),
+
+  // --- LLM audit trail (Bug8) ------------------------------------------------
+  // Records every background LLM call (autoDream consolidation + summary,
+  // autoSummarize compression) into llm_audit_logs: tokens, duration, status
+  // and which trigger produced it. Failures are recorded as status=error and
+  // never block the feature. retentionDays bounds the table: older rows are
+  // purged on boot.
+  llmAudit: z.object({
+    enabled: z.boolean().default(true),
+    retentionDays: z.natural().min(1).max(3650).default(90)
+  }).default({}),
+
   // --- recall evaluation: test-result storage (v0.4.6, 方案 B) --------------
   // Separate retrieval evaluation snapshots from the production recall audit.
   // When false (default) evaluateRetrieval still computes precision/recall/mrr