memorix 1.1.7 → 1.1.8

package/src/llm/memory-manager.ts

@@ -1,328 +1,328 @@
-/**
- * Memory Compact Engine
- *
- * Dual-mode memory management inspired by:
- * - Mem0's ADD/UPDATE/DELETE/NONE decision model with content merging
- * - Cipher's dual-mode architecture (LLM + heuristic fallback)
- *
- * Two paths:
- * 1. LLM mode (compactOnWrite): Single LLM call → extract facts + compare + decide + merge
- * 2. Free mode (heuristicCompact): Vector similarity → rule-based ADD/UPDATE/NONE
- *
- * Both paths produce a CompactDecision that the caller executes.
- */
-
-import { callLLM, isLLMEnabled } from './provider.js';
-
-/** The decision the compact engine makes for each memory operation */
-export type MemoryAction = 'ADD' | 'UPDATE' | 'DELETE' | 'NONE';
-
-/** Existing memory entry for comparison */
-export interface ExistingMemory {
-  id: number;
-  title: string;
-  narrative: string;
-  facts: string;
-  score: number;
-}
-
-/** The compact decision — what to do with the new memory */
-export interface CompactDecision {
-  action: MemoryAction;
-  /** ID of existing memory to UPDATE/DELETE */
-  targetId?: number;
-  /** Brief explanation of why this action was chosen */
-  reason: string;
-  /** Merged narrative for UPDATE (Mem0-style rewrite) */
-  mergedNarrative?: string;
-  /** Merged facts for UPDATE */
-  mergedFacts?: string[];
-  /** LLM-extracted enriched facts (only in LLM mode) */
-  enrichedFacts?: string[];
-  /** Whether LLM was used for this decision */
-  usedLLM: boolean;
-}
-
-/**
- * Unified Compact on Write prompt (Mem0-inspired, single LLM call).
- *
- * This prompt does 3 things in 1 call:
- * 1. Extract structured facts from the new content
- * 2. Compare with existing similar memories
- * 3. Decide ADD/UPDATE/DELETE/NONE and merge if needed
- */
-const COMPACT_ON_WRITE_PROMPT = `You are a smart coding memory manager. You control the memory of a cross-IDE coding assistant.
-
-You receive a NEW MEMORY to store and a list of EXISTING similar memories. Your job:
-1. Extract the key facts from the new memory
-2. Compare with existing memories
-3. Decide the best action
-
-Actions:
-- ADD: New memory contains unique information. Store it as-is.
-- UPDATE: New memory supersedes or improves an existing one. Merge them into a single, comprehensive memory.
-- DELETE: An existing memory is outdated/contradicted by the new one. Remove it.
-- NONE: New memory is redundant. Existing memories already cover this. Skip storing.
-
-Decision rules:
-- Same topic updated (e.g., "MySQL → PostgreSQL"): UPDATE the old memory with merged content
-- Bug fixed that was reported as open: UPDATE the bug report to include the fix
-- Task completed that was tracked as in-progress: UPDATE to mark completed
-- Minor variation of existing memory: NONE (skip)
-- Completely new topic: ADD
-- Old info directly contradicted: DELETE the old one
-- Prefer UPDATE over ADD — keep memory count low, merge information
-
-For UPDATE: write a merged narrative that combines the best of both old and new, preserving all important details. Also merge the facts lists, removing duplicates.
-
-Respond in JSON only:
-{
-  "action": "ADD" | "UPDATE" | "DELETE" | "NONE",
-  "targetId": null or existing_memory_id_number,
-  "reason": "brief explanation of decision",
-  "mergedNarrative": "merged narrative text (required for UPDATE, null otherwise)",
-  "mergedFacts": ["merged fact 1", "merged fact 2"] or null,
-  "extractedFacts": ["fact extracted from new content 1", "fact 2"]
-}`;
-
-// ─── Heuristic Constants (Cipher-inspired) ───────────────────────────
-
-/** Similarity threshold above which we consider memories as "same topic" */
-const SIMILARITY_HIGH = 0.75;
-/** Similarity threshold for "related but different" */
-const SIMILARITY_MEDIUM = 0.5;
-
-// ─── LLM Mode ────────────────────────────────────────────────────────
-
-/**
- * Compact on Write — LLM mode.
- *
- * Single LLM call that extracts facts, compares with existing memories,
- * and decides ADD/UPDATE/DELETE/NONE with merged content.
- *
- * Inspired by Mem0's `get_update_memory_messages` but unified into 1 call.
- */
-export async function compactOnWrite(
-  newMemory: { title: string; narrative: string; facts: string[] },
-  existingMemories: ExistingMemory[],
-): Promise<CompactDecision> {
-  if (!isLLMEnabled()) {
-    return heuristicCompact(newMemory, existingMemories);
-  }
-
-  if (existingMemories.length === 0) {
-    return { action: 'ADD', reason: 'No existing memories to compare', usedLLM: false };
-  }
-
-  // Build the prompt with new + existing memories (Mem0's temp_uuid_mapping approach)
-  const existingList = existingMemories
-    .map(m => `[ID: ${m.id}] (similarity: ${m.score.toFixed(2)}) ${m.title}\n  Narrative: ${m.narrative}\n  Facts: ${m.facts}`)
-    .join('\n\n');
-
-  const userMessage = `NEW MEMORY:
-Title: ${newMemory.title}
-Narrative: ${newMemory.narrative}
-Facts: ${newMemory.facts.join('; ')}
-
-EXISTING SIMILAR MEMORIES:
-${existingList}`;
-
-  try {
-    const response = await callLLM(COMPACT_ON_WRITE_PROMPT, userMessage);
-
-    // Parse response — handle markdown code blocks
-    let content = response.content.trim();
-    if (content.startsWith('```')) {
-      content = content.replace(/^```(?:json)?\s*/, '').replace(/\s*```$/, '');
-    }
-
-    const parsed = JSON.parse(content) as {
-      action: string;
-      targetId?: number | null;
-      reason?: string;
-      mergedNarrative?: string | null;
-      mergedFacts?: string[] | null;
-      extractedFacts?: string[] | null;
-    };
-
-    // Validate action
-    const action = parsed.action?.toUpperCase() as MemoryAction;
-    if (!action || !['ADD', 'UPDATE', 'DELETE', 'NONE'].includes(action)) {
-      return { action: 'ADD', reason: 'LLM response invalid, defaulting to ADD', usedLLM: true };
-    }
-
-    // Validate targetId exists in existing memories for UPDATE/DELETE
-    if ((action === 'UPDATE' || action === 'DELETE') && parsed.targetId != null) {
-      const targetExists = existingMemories.some(m => m.id === parsed.targetId);
-      if (!targetExists) {
-        // LLM hallucinated an ID (common issue noted in Mem0's code)
-        return { action: 'ADD', reason: `LLM referenced non-existent memory #${parsed.targetId}, defaulting to ADD`, usedLLM: true };
-      }
-    }
-
-    return {
-      action,
-      targetId: parsed.targetId ?? undefined,
-      reason: parsed.reason ?? 'LLM decision',
-      mergedNarrative: parsed.mergedNarrative ?? undefined,
-      mergedFacts: parsed.mergedFacts ?? undefined,
-      enrichedFacts: parsed.extractedFacts ?? undefined,
-      usedLLM: true,
-    };
-  } catch (err) {
-    // LLM failed — fall back to heuristic
-    console.error(`[memorix] LLM compact failed, falling back to heuristic:`, (err as Error)?.message ?? err);
-    return heuristicCompact(newMemory, existingMemories);
-  }
-}
-
-// ─── Free Mode (Heuristic) ───────────────────────────────────────────
-
-/**
- * Heuristic Compact — Free mode, no LLM needed.
- *
- * Uses vector similarity scores from search results to make decisions.
- * Inspired by Cipher's fallback logic in extract_and_operate_memory.ts.
- *
- * Decision logic:
- * - score >= 0.75: Very similar → check if new is more complete
- *   - New is longer/richer → UPDATE (merge)
- *   - Otherwise → NONE (skip, already covered)
- * - score >= 0.50: Related topic
- *   - Same entity + same type → UPDATE if new has more facts
- *   - Otherwise → ADD (different enough)
- * - score < 0.50: Different topic → ADD
- */
-export function heuristicCompact(
-  newMemory: { title: string; narrative: string; facts: string[] },
-  existingMemories: ExistingMemory[],
-): CompactDecision {
-  if (existingMemories.length === 0) {
-    return { action: 'ADD', reason: 'No existing memories to compare', usedLLM: false };
-  }
-
-  const best = existingMemories[0]; // Already sorted by score
-
-  // High similarity — very likely same topic
-  if (best.score >= SIMILARITY_HIGH) {
-    const newLength = newMemory.narrative.length + newMemory.facts.join(' ').length;
-    const oldLength = best.narrative.length + best.facts.length;
-
-    if (newLength > oldLength * 1.2) {
-      // New memory is substantially richer — UPDATE with merged content
-      const mergedNarrative = mergeTexts(best.narrative, newMemory.narrative);
-      const mergedFacts = mergeFacts(best.facts, newMemory.facts.join('\n'));
-      return {
-        action: 'UPDATE',
-        targetId: best.id,
-        reason: `New memory is more comprehensive (${newLength} > ${oldLength} chars)`,
-        mergedNarrative,
-        mergedFacts,
-        usedLLM: false,
-      };
-    }
-
-    // New memory is not richer — skip it
-    return {
-      action: 'NONE',
-      targetId: best.id,
-      reason: `Existing memory #${best.id} already covers this topic (similarity: ${best.score.toFixed(2)})`,
-      usedLLM: false,
-    };
-  }
-
-  // Medium similarity — related but might be different enough
-  if (best.score >= SIMILARITY_MEDIUM) {
-    const newFactCount = newMemory.facts.length;
-    const oldFactCount = best.facts.split('\n').filter(Boolean).length;
-
-    if (newFactCount > oldFactCount) {
-      // New has more facts — UPDATE
-      const mergedNarrative = mergeTexts(best.narrative, newMemory.narrative);
-      const mergedFacts = mergeFacts(best.facts, newMemory.facts.join('\n'));
-      return {
-        action: 'UPDATE',
-        targetId: best.id,
-        reason: `New memory has more facts (${newFactCount} > ${oldFactCount}), merging`,
-        mergedNarrative,
-        mergedFacts,
-        usedLLM: false,
-      };
-    }
-
-    // Related but not richer — ADD as separate memory
-    return { action: 'ADD', reason: `Related to #${best.id} but different enough to keep separate`, usedLLM: false };
-  }
-
-  // Low similarity — new topic
-  return { action: 'ADD', reason: 'No similar memories found', usedLLM: false };
-}
-
-// ─── Content Merging Utilities ───────────────────────────────────────
-
-/**
- * Merge two narrative texts, keeping the more comprehensive version
- * while preserving unique information from both.
- */
-function mergeTexts(oldText: string, newText: string): string {
-  // If new text is significantly longer, prefer it
-  if (newText.length > oldText.length * 1.5) return newText;
-  // If old text is significantly longer, append new info
-  if (oldText.length > newText.length * 1.5) return oldText;
-  // Similar length — combine with separator
-  return `${newText}\n\n[Previous context]: ${oldText}`;
-}
-
-/**
- * Merge two fact lists, removing duplicates.
- * oldFacts is newline-separated string, newFacts is array.
- */
-function mergeFacts(oldFactsStr: string, newFactsStr: string): string[] {
-  const oldFacts = oldFactsStr.split('\n').filter(Boolean).map(f => f.trim());
-  const newFacts = newFactsStr.split('\n').filter(Boolean).map(f => f.trim());
-
-  const seen = new Set<string>();
-  const merged: string[] = [];
-
-  // Add new facts first (they're more recent)
-  for (const f of newFacts) {
-    const normalized = f.toLowerCase();
-    if (!seen.has(normalized)) {
-      seen.add(normalized);
-      merged.push(f);
-    }
-  }
-
-  // Add old facts that aren't duplicates
-  for (const f of oldFacts) {
-    const normalized = f.toLowerCase();
-    if (!seen.has(normalized)) {
-      seen.add(normalized);
-      merged.push(f);
-    }
-  }
-
-  return merged;
-}
-
-// ─── Batch Dedup (for memorix_deduplicate tool) ──────────────────────
-
-/**
- * LLM-powered dedup for batch cleanup.
- * Compares a new memory against existing ones and returns a decision.
- * Used by memorix_deduplicate tool.
- */
-export async function deduplicateMemory(
-  newMemory: { title: string; narrative: string; facts: string[] },
-  existingMemories: Array<{ id: number; title: string; narrative: string; facts: string }>,
-): Promise<CompactDecision | null> {
-  if (!isLLMEnabled()) return null;
-  if (existingMemories.length === 0) return { action: 'ADD', reason: 'No existing memories', usedLLM: false };
-
-  const asExisting: ExistingMemory[] = existingMemories.map(m => ({
-    ...m,
-    score: 0.8, // Batch dedup assumes high similarity (pre-filtered)
-  }));
-
-  return compactOnWrite(newMemory, asExisting);
-}
+/**
+ * Memory Compact Engine
+ *
+ * Dual-mode memory management inspired by:
+ * - Mem0's ADD/UPDATE/DELETE/NONE decision model with content merging
+ * - Cipher's dual-mode architecture (LLM + heuristic fallback)
+ *
+ * Two paths:
+ * 1. LLM mode (compactOnWrite): Single LLM call → extract facts + compare + decide + merge
+ * 2. Free mode (heuristicCompact): Vector similarity → rule-based ADD/UPDATE/NONE
+ *
+ * Both paths produce a CompactDecision that the caller executes.
+ */
+
+import { callLLM, isLLMEnabled } from './provider.js';
+
+/** The decision the compact engine makes for each memory operation */
+export type MemoryAction = 'ADD' | 'UPDATE' | 'DELETE' | 'NONE';
+
+/** Existing memory entry for comparison */
+export interface ExistingMemory {
+  id: number;
+  title: string;
+  narrative: string;
+  facts: string;
+  score: number;
+}
+
+/** The compact decision — what to do with the new memory */
+export interface CompactDecision {
+  action: MemoryAction;
+  /** ID of existing memory to UPDATE/DELETE */
+  targetId?: number;
+  /** Brief explanation of why this action was chosen */
+  reason: string;
+  /** Merged narrative for UPDATE (Mem0-style rewrite) */
+  mergedNarrative?: string;
+  /** Merged facts for UPDATE */
+  mergedFacts?: string[];
+  /** LLM-extracted enriched facts (only in LLM mode) */
+  enrichedFacts?: string[];
+  /** Whether LLM was used for this decision */
+  usedLLM: boolean;
+}
+
+/**
+ * Unified Compact on Write prompt (Mem0-inspired, single LLM call).
+ *
+ * This prompt does 3 things in 1 call:
+ * 1. Extract structured facts from the new content
+ * 2. Compare with existing similar memories
+ * 3. Decide ADD/UPDATE/DELETE/NONE and merge if needed
+ */
+const COMPACT_ON_WRITE_PROMPT = `You are a smart coding memory manager. You control the memory of a cross-IDE coding assistant.
+
+You receive a NEW MEMORY to store and a list of EXISTING similar memories. Your job:
+1. Extract the key facts from the new memory
+2. Compare with existing memories
+3. Decide the best action
+
+Actions:
+- ADD: New memory contains unique information. Store it as-is.
+- UPDATE: New memory supersedes or improves an existing one. Merge them into a single, comprehensive memory.
+- DELETE: An existing memory is outdated/contradicted by the new one. Remove it.
+- NONE: New memory is redundant. Existing memories already cover this. Skip storing.
+
+Decision rules:
+- Same topic updated (e.g., "MySQL → PostgreSQL"): UPDATE the old memory with merged content
+- Bug fixed that was reported as open: UPDATE the bug report to include the fix
+- Task completed that was tracked as in-progress: UPDATE to mark completed
+- Minor variation of existing memory: NONE (skip)
+- Completely new topic: ADD
+- Old info directly contradicted: DELETE the old one
+- Prefer UPDATE over ADD — keep memory count low, merge information
+
+For UPDATE: write a merged narrative that combines the best of both old and new, preserving all important details. Also merge the facts lists, removing duplicates.
+
+Respond in JSON only:
+{
+  "action": "ADD" | "UPDATE" | "DELETE" | "NONE",
+  "targetId": null or existing_memory_id_number,
+  "reason": "brief explanation of decision",
+  "mergedNarrative": "merged narrative text (required for UPDATE, null otherwise)",
+  "mergedFacts": ["merged fact 1", "merged fact 2"] or null,
+  "extractedFacts": ["fact extracted from new content 1", "fact 2"]
+}`;
+
+// ─── Heuristic Constants (Cipher-inspired) ───────────────────────────
+
+/** Similarity threshold above which we consider memories as "same topic" */
+const SIMILARITY_HIGH = 0.75;
+/** Similarity threshold for "related but different" */
+const SIMILARITY_MEDIUM = 0.5;
+
+// ─── LLM Mode ────────────────────────────────────────────────────────
+
+/**
+ * Compact on Write — LLM mode.
+ *
+ * Single LLM call that extracts facts, compares with existing memories,
+ * and decides ADD/UPDATE/DELETE/NONE with merged content.
+ *
+ * Inspired by Mem0's `get_update_memory_messages` but unified into 1 call.
+ */
+export async function compactOnWrite(
+  newMemory: { title: string; narrative: string; facts: string[] },
+  existingMemories: ExistingMemory[],
+): Promise<CompactDecision> {
+  if (!isLLMEnabled()) {
+    return heuristicCompact(newMemory, existingMemories);
+  }
+
+  if (existingMemories.length === 0) {
+    return { action: 'ADD', reason: 'No existing memories to compare', usedLLM: false };
+  }
+
+  // Build the prompt with new + existing memories (Mem0's temp_uuid_mapping approach)
+  const existingList = existingMemories
+    .map(m => `[ID: ${m.id}] (similarity: ${m.score.toFixed(2)}) ${m.title}\n  Narrative: ${m.narrative}\n  Facts: ${m.facts}`)
+    .join('\n\n');
+
+  const userMessage = `NEW MEMORY:
+Title: ${newMemory.title}
+Narrative: ${newMemory.narrative}
+Facts: ${newMemory.facts.join('; ')}
+
+EXISTING SIMILAR MEMORIES:
+${existingList}`;
+
+  try {
+    const response = await callLLM(COMPACT_ON_WRITE_PROMPT, userMessage);
+
+    // Parse response — handle markdown code blocks
+    let content = response.content.trim();
+    if (content.startsWith('```')) {
+      content = content.replace(/^```(?:json)?\s*/, '').replace(/\s*```$/, '');
+    }
+
+    const parsed = JSON.parse(content) as {
+      action: string;
+      targetId?: number | null;
+      reason?: string;
+      mergedNarrative?: string | null;
+      mergedFacts?: string[] | null;
+      extractedFacts?: string[] | null;
+    };
+
+    // Validate action
+    const action = parsed.action?.toUpperCase() as MemoryAction;
+    if (!action || !['ADD', 'UPDATE', 'DELETE', 'NONE'].includes(action)) {
+      return { action: 'ADD', reason: 'LLM response invalid, defaulting to ADD', usedLLM: true };
+    }
+
+    // Validate targetId exists in existing memories for UPDATE/DELETE
+    if ((action === 'UPDATE' || action === 'DELETE') && parsed.targetId != null) {
+      const targetExists = existingMemories.some(m => m.id === parsed.targetId);
+      if (!targetExists) {
+        // LLM hallucinated an ID (common issue noted in Mem0's code)
+        return { action: 'ADD', reason: `LLM referenced non-existent memory #${parsed.targetId}, defaulting to ADD`, usedLLM: true };
+      }
+    }
+
+    return {
+      action,
+      targetId: parsed.targetId ?? undefined,
+      reason: parsed.reason ?? 'LLM decision',
+      mergedNarrative: parsed.mergedNarrative ?? undefined,
+      mergedFacts: parsed.mergedFacts ?? undefined,
+      enrichedFacts: parsed.extractedFacts ?? undefined,
+      usedLLM: true,
+    };
+  } catch (err) {
+    // LLM failed — fall back to heuristic
+    console.error(`[memorix] LLM compact failed, falling back to heuristic:`, (err as Error)?.message ?? err);
+    return heuristicCompact(newMemory, existingMemories);
+  }
+}
+
+// ─── Free Mode (Heuristic) ───────────────────────────────────────────
+
+/**
+ * Heuristic Compact — Free mode, no LLM needed.
+ *
+ * Uses vector similarity scores from search results to make decisions.
+ * Inspired by Cipher's fallback logic in extract_and_operate_memory.ts.
+ *
+ * Decision logic:
+ * - score >= 0.75: Very similar → check if new is more complete
+ *   - New is longer/richer → UPDATE (merge)
+ *   - Otherwise → NONE (skip, already covered)
+ * - score >= 0.50: Related topic
+ *   - Same entity + same type → UPDATE if new has more facts
+ *   - Otherwise → ADD (different enough)
+ * - score < 0.50: Different topic → ADD
+ */
+export function heuristicCompact(
+  newMemory: { title: string; narrative: string; facts: string[] },
+  existingMemories: ExistingMemory[],
+): CompactDecision {
+  if (existingMemories.length === 0) {
+    return { action: 'ADD', reason: 'No existing memories to compare', usedLLM: false };
+  }
+
+  const best = existingMemories[0]; // Already sorted by score
+
+  // High similarity — very likely same topic
+  if (best.score >= SIMILARITY_HIGH) {
+    const newLength = newMemory.narrative.length + newMemory.facts.join(' ').length;
+    const oldLength = best.narrative.length + best.facts.length;
+
+    if (newLength > oldLength * 1.2) {
+      // New memory is substantially richer — UPDATE with merged content
+      const mergedNarrative = mergeTexts(best.narrative, newMemory.narrative);
+      const mergedFacts = mergeFacts(best.facts, newMemory.facts.join('\n'));
+      return {
+        action: 'UPDATE',
+        targetId: best.id,
+        reason: `New memory is more comprehensive (${newLength} > ${oldLength} chars)`,
+        mergedNarrative,
+        mergedFacts,
+        usedLLM: false,
+      };
+    }
+
+    // New memory is not richer — skip it
+    return {
+      action: 'NONE',
+      targetId: best.id,
+      reason: `Existing memory #${best.id} already covers this topic (similarity: ${best.score.toFixed(2)})`,
+      usedLLM: false,
+    };
+  }
+
+  // Medium similarity — related but might be different enough
+  if (best.score >= SIMILARITY_MEDIUM) {
+    const newFactCount = newMemory.facts.length;
+    const oldFactCount = best.facts.split('\n').filter(Boolean).length;
+
+    if (newFactCount > oldFactCount) {
+      // New has more facts — UPDATE
+      const mergedNarrative = mergeTexts(best.narrative, newMemory.narrative);
+      const mergedFacts = mergeFacts(best.facts, newMemory.facts.join('\n'));
+      return {
+        action: 'UPDATE',
+        targetId: best.id,
+        reason: `New memory has more facts (${newFactCount} > ${oldFactCount}), merging`,
+        mergedNarrative,
+        mergedFacts,
+        usedLLM: false,
+      };
+    }
+
+    // Related but not richer — ADD as separate memory
+    return { action: 'ADD', reason: `Related to #${best.id} but different enough to keep separate`, usedLLM: false };
+  }
+
+  // Low similarity — new topic
+  return { action: 'ADD', reason: 'No similar memories found', usedLLM: false };
+}
+
+// ─── Content Merging Utilities ───────────────────────────────────────
+
+/**
+ * Merge two narrative texts, keeping the more comprehensive version
+ * while preserving unique information from both.
+ */
+function mergeTexts(oldText: string, newText: string): string {
+  // If new text is significantly longer, prefer it
+  if (newText.length > oldText.length * 1.5) return newText;
+  // If old text is significantly longer, append new info
+  if (oldText.length > newText.length * 1.5) return oldText;
+  // Similar length — combine with separator
+  return `${newText}\n\n[Previous context]: ${oldText}`;
+}
+
+/**
+ * Merge two fact lists, removing duplicates.
+ * oldFacts is newline-separated string, newFacts is array.
+ */
+function mergeFacts(oldFactsStr: string, newFactsStr: string): string[] {
+  const oldFacts = oldFactsStr.split('\n').filter(Boolean).map(f => f.trim());
+  const newFacts = newFactsStr.split('\n').filter(Boolean).map(f => f.trim());
+
+  const seen = new Set<string>();
+  const merged: string[] = [];
+
+  // Add new facts first (they're more recent)
+  for (const f of newFacts) {
+    const normalized = f.toLowerCase();
+    if (!seen.has(normalized)) {
+      seen.add(normalized);
+      merged.push(f);
+    }
+  }
+
+  // Add old facts that aren't duplicates
+  for (const f of oldFacts) {
+    const normalized = f.toLowerCase();
+    if (!seen.has(normalized)) {
+      seen.add(normalized);
+      merged.push(f);
+    }
+  }
+
+  return merged;
+}
+
+// ─── Batch Dedup (for memorix_deduplicate tool) ──────────────────────
+
+/**
+ * LLM-powered dedup for batch cleanup.
+ * Compares a new memory against existing ones and returns a decision.
+ * Used by memorix_deduplicate tool.
+ */
+export async function deduplicateMemory(
+  newMemory: { title: string; narrative: string; facts: string[] },
+  existingMemories: Array<{ id: number; title: string; narrative: string; facts: string }>,
+): Promise<CompactDecision | null> {
+  if (!isLLMEnabled()) return null;
+  if (existingMemories.length === 0) return { action: 'ADD', reason: 'No existing memories', usedLLM: false };
+
+  const asExisting: ExistingMemory[] = existingMemories.map(m => ({
+    ...m,
+    score: 0.8, // Batch dedup assumes high similarity (pre-filtered)
+  }));
+
+  return compactOnWrite(newMemory, asExisting);
+}