memorix 1.1.7 → 1.1.9

package/src/memory/formation/resolve.ts

@@ -1,82 +1,82 @@
-/**
- * Memory Formation — Stage 2: Resolve
- *
- * Determines what to do with an enriched memory: create new, merge into
- * existing, evolve (supersede), or discard as redundant.
- *
- * This stage absorbs and replaces the previous "Compact on Write" logic
- * (src/llm/memory-manager.ts) with a richer resolution model:
- *
- * - new:     Truly new knowledge → proceed to store
- * - merge:   Same topic as existing → UPDATE with combined content
- * - evolve:  Existing is outdated → UPDATE with new content as primary
- * - discard: Redundant or noise → skip storage entirely
- *
- * Rules-based mode uses similarity scores, entity overlap, fact comparison,
- * and contradiction detection to make decisions without LLM.
- */
-
-import type { ExtractResult, ResolveResult, SearchHit, ExistingMemoryRef } from './types.js';
-
-// ── Thresholds ───────────────────────────────────────────────────
-
-/** Above this: very likely same topic */
-const SIMILARITY_HIGH = 0.75;
-/** Above this: related topic */
-const SIMILARITY_MEDIUM = 0.50;
-/** Above this: exact duplicate — discard */
-const SIMILARITY_DUPLICATE = 0.90;
-
-// ── Content Comparison Utilities ─────────────────────────────────
-
-/**
- * Compute Jaccard similarity between two sets of normalized words.
- */
-function wordOverlap(a: string, b: string): number {
-  const wordsA = new Set(a.toLowerCase().split(/\s+/).filter(w => w.length > 2));
-  const wordsB = new Set(b.toLowerCase().split(/\s+/).filter(w => w.length > 2));
-  if (wordsA.size === 0 || wordsB.size === 0) return 0;
-
-  let intersection = 0;
-  for (const w of wordsA) {
-    if (wordsB.has(w)) intersection++;
-  }
-  return intersection / Math.max(wordsA.size, wordsB.size);
-}
-
-/**
- * Check if two entity names refer to the same concept.
- */
-function entitiesMatch(a: string, b: string): boolean {
-  const na = a.toLowerCase().replace(/[-_]/g, '');
-  const nb = b.toLowerCase().replace(/[-_]/g, '');
-  if (na === nb) return true;
-  if (na.length >= 3 && nb.length >= 3) {
-    if (na.includes(nb) || nb.includes(na)) return true;
-  }
-  return false;
-}
-
-/**
- * Detect potential contradiction between old and new content.
- * Looks for negation patterns and opposing statements.
- */
+/**
+ * Memory Formation — Stage 2: Resolve
+ *
+ * Determines what to do with an enriched memory: create new, merge into
+ * existing, evolve (supersede), or discard as redundant.
+ *
+ * This stage absorbs and replaces the previous "Compact on Write" logic
+ * (src/llm/memory-manager.ts) with a richer resolution model:
+ *
+ * - new:     Truly new knowledge → proceed to store
+ * - merge:   Same topic as existing → UPDATE with combined content
+ * - evolve:  Existing is outdated → UPDATE with new content as primary
+ * - discard: Redundant or noise → skip storage entirely
+ *
+ * Rules-based mode uses similarity scores, entity overlap, fact comparison,
+ * and contradiction detection to make decisions without LLM.
+ */
+
+import type { ExtractResult, ResolveResult, SearchHit, ExistingMemoryRef } from './types.js';
+
+// ── Thresholds ───────────────────────────────────────────────────
+
+/** Above this: very likely same topic */
+const SIMILARITY_HIGH = 0.75;
+/** Above this: related topic */
+const SIMILARITY_MEDIUM = 0.50;
+/** Above this: exact duplicate — discard */
+const SIMILARITY_DUPLICATE = 0.90;
+
+// ── Content Comparison Utilities ─────────────────────────────────
+
+/**
+ * Compute Jaccard similarity between two sets of normalized words.
+ */
+function wordOverlap(a: string, b: string): number {
+  const wordsA = new Set(a.toLowerCase().split(/\s+/).filter(w => w.length > 2));
+  const wordsB = new Set(b.toLowerCase().split(/\s+/).filter(w => w.length > 2));
+  if (wordsA.size === 0 || wordsB.size === 0) return 0;
+
+  let intersection = 0;
+  for (const w of wordsA) {
+    if (wordsB.has(w)) intersection++;
+  }
+  return intersection / Math.max(wordsA.size, wordsB.size);
+}
+
+/**
+ * Check if two entity names refer to the same concept.
+ */
+function entitiesMatch(a: string, b: string): boolean {
+  const na = a.toLowerCase().replace(/[-_]/g, '');
+  const nb = b.toLowerCase().replace(/[-_]/g, '');
+  if (na === nb) return true;
+  if (na.length >= 3 && nb.length >= 3) {
+    if (na.includes(nb) || nb.includes(na)) return true;
+  }
+  return false;
+}
+
+/**
+ * Detect potential contradiction between old and new content.
+ * Looks for negation patterns and opposing statements.
+ */
 function hasContradiction(oldText: string, newText: string): boolean {
-  // Simple heuristic: check for "not X" in new when "X" is in old
-  const negationPatterns = [
-    /\bnot\s+(\w+)/gi,
-    /\bno longer\b/i,
-    /\binstead of\b/i,
-    /\breplaced\b.*\bwith\b/i,
-    /\bremoved\b/i,
-    /\bdeprecated\b/i,
-    /\bobsolete\b/i,
-    /不再/,
-    /已弃用/,
-    /替换为/,
-    /改为/,
-  ];
-
+  // Simple heuristic: check for "not X" in new when "X" is in old
+  const negationPatterns = [
+    /\bnot\s+(\w+)/gi,
+    /\bno longer\b/i,
+    /\binstead of\b/i,
+    /\breplaced\b.*\bwith\b/i,
+    /\bremoved\b/i,
+    /\bdeprecated\b/i,
+    /\bobsolete\b/i,
+    /不再/,
+    /已弃用/,
+    /替换为/,
+    /改为/,
+  ];
+
   return negationPatterns.some(p => p.test(newText));
 }
 
@@ -90,141 +90,141 @@ function normalizedSearchSimilarity(score: number): number {
   if (score <= 1) return score;
   return 0;
 }
-
-/**
- * Merge two narratives, keeping the most comprehensive version.
- */
-function mergeNarratives(oldNarrative: string, newNarrative: string): string {
-  if (newNarrative.length > oldNarrative.length * 1.5) return newNarrative;
-  if (oldNarrative.length > newNarrative.length * 1.5) return oldNarrative;
-  return `${newNarrative}\n\n[Previous context]: ${oldNarrative}`;
-}
-
-/**
- * Merge two fact lists, deduplicating by normalized text.
- */
-function mergeFacts(oldFacts: string[], newFacts: string[]): string[] {
-  const seen = new Set<string>();
-  const merged: string[] = [];
-
-  // New facts first (more recent)
-  for (const f of newFacts) {
-    const norm = f.toLowerCase().trim();
-    if (!seen.has(norm) && f.trim().length > 0) {
-      seen.add(norm);
-      merged.push(f);
-    }
-  }
-  for (const f of oldFacts) {
-    const norm = f.toLowerCase().trim();
-    if (!seen.has(norm) && f.trim().length > 0) {
-      seen.add(norm);
-      merged.push(f);
-    }
-  }
-
-  return merged;
-}
-
-// ── LLM Resolution (Mem0-style) ───────────────────────────────
-
-const LLM_RESOLVE_PROMPT = `You are a Memory Consolidation Manager for a software engineering knowledge base.
-
-You must decide what to do with a NEW memory given EXISTING memories that are similar.
-
-Operations:
-- ADD: The new memory contains genuinely new information not present in existing memories.
-- UPDATE: The new memory adds to or refines an existing memory. Specify which existing memory ID to update.
-- DELETE: The new memory contradicts an existing memory. Specify which existing memory ID to delete.
-- NOOP: The new memory is redundant (already covered by existing memories). Skip storage.
-
-Rules:
-- Return ONLY a JSON object
-- If UPDATE: merge the best information from both old and new
-- If DELETE: the new memory supersedes the old (contradiction detected)
-- Prefer UPDATE over ADD when the topic is the same but information differs
-- Prefer NOOP over ADD when the information is essentially the same
-
-Response format:
-{"action": "ADD|UPDATE|DELETE|NOOP", "targetId": <number or null>, "reason": "<brief explanation>", "mergedText": "<merged content for UPDATE, or null>"}`;
-
-async function resolveWithLLM(
-  extracted: ExtractResult,
-  hits: SearchHit[],
-  getObservation: (id: number) => ExistingMemoryRef | null,
-): Promise<ResolveResult | null> {
-  try {
-    const { callLLM } = await import('../../llm/provider.js');
-
-    // Build context for LLM
-    const existingMemories = hits.slice(0, 5).map((h, i) => ({
-      id: h.observationId,
-      index: i,
-      title: h.title,
-      content: h.narrative.substring(0, 300),
-      facts: h.facts.substring(0, 200),
-    }));
-
-    const input = `NEW MEMORY:
-Title: ${extracted.title}
-Content: ${extracted.narrative.substring(0, 500)}
-Facts: ${extracted.facts.join('; ')}
-
-EXISTING MEMORIES:
-${existingMemories.map(m => `[ID:${m.id}] ${m.title} | ${m.content} | Facts: ${m.facts}`).join('\n')}`;
-
-    const response = await callLLM(LLM_RESOLVE_PROMPT, input);
-    const text = response.content.trim();
-
-    const jsonMatch = text.match(/\{[\s\S]*\}/);
-    if (!jsonMatch) return null;
-    const parsed = JSON.parse(jsonMatch[0]);
-
-    const action = parsed.action?.toUpperCase();
-    const targetId = parsed.targetId ? Number(parsed.targetId) : undefined;
-    const reason = parsed.reason || 'LLM decision';
-
-    if (action === 'NOOP') {
-      return { action: 'discard', targetId, reason: `LLM: ${reason}` };
-    }
-    if (action === 'ADD') {
-      return { action: 'new', reason: `LLM: ${reason}` };
-    }
-    if (action === 'UPDATE' && targetId) {
-      const existing = getObservation(targetId);
-      const oldFacts = existing?.facts ?? [];
-      return {
-        action: 'merge',
-        targetId,
-        reason: `LLM: ${reason}`,
-        mergedNarrative: parsed.mergedText || mergeNarratives(existing?.narrative ?? '', extracted.narrative),
-        mergedFacts: mergeFacts(oldFacts, extracted.facts),
-      };
-    }
-    if (action === 'DELETE' && targetId) {
-      const existing = getObservation(targetId);
-      const oldFacts = existing?.facts ?? [];
-      return {
-        action: 'evolve',
-        targetId,
-        reason: `LLM: ${reason}`,
-        mergedNarrative: extracted.narrative,
-        mergedFacts: mergeFacts(oldFacts, extracted.facts),
-      };
-    }
-
-    return null; // Unrecognized action
-  } catch {
-    return null; // LLM failure → fall back to rules
-  }
-}
-
-// ── Resolve Implementation ───────────────────────────────────────
-
-/**
- * Score a candidate match for resolution.
- * Returns a composite score considering similarity, entity overlap, and content richness.
- */
+
+/**
+ * Merge two narratives, keeping the most comprehensive version.
+ */
+function mergeNarratives(oldNarrative: string, newNarrative: string): string {
+  if (newNarrative.length > oldNarrative.length * 1.5) return newNarrative;
+  if (oldNarrative.length > newNarrative.length * 1.5) return oldNarrative;
+  return `${newNarrative}\n\n[Previous context]: ${oldNarrative}`;
+}
+
+/**
+ * Merge two fact lists, deduplicating by normalized text.
+ */
+function mergeFacts(oldFacts: string[], newFacts: string[]): string[] {
+  const seen = new Set<string>();
+  const merged: string[] = [];
+
+  // New facts first (more recent)
+  for (const f of newFacts) {
+    const norm = f.toLowerCase().trim();
+    if (!seen.has(norm) && f.trim().length > 0) {
+      seen.add(norm);
+      merged.push(f);
+    }
+  }
+  for (const f of oldFacts) {
+    const norm = f.toLowerCase().trim();
+    if (!seen.has(norm) && f.trim().length > 0) {
+      seen.add(norm);
+      merged.push(f);
+    }
+  }
+
+  return merged;
+}
+
+// ── LLM Resolution (Mem0-style) ───────────────────────────────
+
+const LLM_RESOLVE_PROMPT = `You are a Memory Consolidation Manager for a software engineering knowledge base.
+
+You must decide what to do with a NEW memory given EXISTING memories that are similar.
+
+Operations:
+- ADD: The new memory contains genuinely new information not present in existing memories.
+- UPDATE: The new memory adds to or refines an existing memory. Specify which existing memory ID to update.
+- DELETE: The new memory contradicts an existing memory. Specify which existing memory ID to delete.
+- NOOP: The new memory is redundant (already covered by existing memories). Skip storage.
+
+Rules:
+- Return ONLY a JSON object
+- If UPDATE: merge the best information from both old and new
+- If DELETE: the new memory supersedes the old (contradiction detected)
+- Prefer UPDATE over ADD when the topic is the same but information differs
+- Prefer NOOP over ADD when the information is essentially the same
+
+Response format:
+{"action": "ADD|UPDATE|DELETE|NOOP", "targetId": <number or null>, "reason": "<brief explanation>", "mergedText": "<merged content for UPDATE, or null>"}`;
+
+async function resolveWithLLM(
+  extracted: ExtractResult,
+  hits: SearchHit[],
+  getObservation: (id: number) => ExistingMemoryRef | null,
+): Promise<ResolveResult | null> {
+  try {
+    const { callLLM } = await import('../../llm/provider.js');
+
+    // Build context for LLM
+    const existingMemories = hits.slice(0, 5).map((h, i) => ({
+      id: h.observationId,
+      index: i,
+      title: h.title,
+      content: h.narrative.substring(0, 300),
+      facts: h.facts.substring(0, 200),
+    }));
+
+    const input = `NEW MEMORY:
+Title: ${extracted.title}
+Content: ${extracted.narrative.substring(0, 500)}
+Facts: ${extracted.facts.join('; ')}
+
+EXISTING MEMORIES:
+${existingMemories.map(m => `[ID:${m.id}] ${m.title} | ${m.content} | Facts: ${m.facts}`).join('\n')}`;
+
+    const response = await callLLM(LLM_RESOLVE_PROMPT, input);
+    const text = response.content.trim();
+
+    const jsonMatch = text.match(/\{[\s\S]*\}/);
+    if (!jsonMatch) return null;
+    const parsed = JSON.parse(jsonMatch[0]);
+
+    const action = parsed.action?.toUpperCase();
+    const targetId = parsed.targetId ? Number(parsed.targetId) : undefined;
+    const reason = parsed.reason || 'LLM decision';
+
+    if (action === 'NOOP') {
+      return { action: 'discard', targetId, reason: `LLM: ${reason}` };
+    }
+    if (action === 'ADD') {
+      return { action: 'new', reason: `LLM: ${reason}` };
+    }
+    if (action === 'UPDATE' && targetId) {
+      const existing = getObservation(targetId);
+      const oldFacts = existing?.facts ?? [];
+      return {
+        action: 'merge',
+        targetId,
+        reason: `LLM: ${reason}`,
+        mergedNarrative: parsed.mergedText || mergeNarratives(existing?.narrative ?? '', extracted.narrative),
+        mergedFacts: mergeFacts(oldFacts, extracted.facts),
+      };
+    }
+    if (action === 'DELETE' && targetId) {
+      const existing = getObservation(targetId);
+      const oldFacts = existing?.facts ?? [];
+      return {
+        action: 'evolve',
+        targetId,
+        reason: `LLM: ${reason}`,
+        mergedNarrative: extracted.narrative,
+        mergedFacts: mergeFacts(oldFacts, extracted.facts),
+      };
+    }
+
+    return null; // Unrecognized action
+  } catch {
+    return null; // LLM failure → fall back to rules
+  }
+}
+
+// ── Resolve Implementation ───────────────────────────────────────
+
+/**
+ * Score a candidate match for resolution.
+ * Returns a composite score considering similarity, entity overlap, and content richness.
+ */
 function scoreCandidate(
   extracted: ExtractResult,
   candidate: SearchHit,
@@ -240,139 +240,139 @@ function scoreCandidate(
   const score = searchSimilarity * 0.6
     + (entityMatch ? 0.2 : 0)
     + contentOverlap * 0.2;
-
-  // Is new memory richer?
-  const newLength = extracted.narrative.length + extracted.facts.join(' ').length;
-  const oldLength = candidate.narrative.length + candidate.facts.length;
-  const richer = newLength > oldLength * 1.15;
-
-  const contradiction = hasContradiction(candidate.narrative, extracted.narrative);
-
+
+  // Is new memory richer?
+  const newLength = extracted.narrative.length + extracted.facts.join(' ').length;
+  const oldLength = candidate.narrative.length + candidate.facts.length;
+  const richer = newLength > oldLength * 1.15;
+
+  const contradiction = hasContradiction(candidate.narrative, extracted.narrative);
+
   return { score, searchSimilarity, entityMatch, richer, contradiction };
 }
-
-/**
- * Run Stage 2: Resolve.
- *
- * Determines the resolution action for an enriched memory by comparing
- * it against existing memories found via search.
- */
-export async function runResolve(
-  extracted: ExtractResult,
-  projectId: string,
-  searchMemories: (query: string, limit: number, projectId: string) => Promise<SearchHit[]>,
-  getObservation: (id: number) => ExistingMemoryRef | null,
-  useLLM = false,
-): Promise<ResolveResult> {
-  // Search for similar existing memories
-  const query = `${extracted.title} ${extracted.narrative.substring(0, 200)}`;
-  let hits: SearchHit[];
-  try {
-    hits = await searchMemories(query, 5, projectId);
-  } catch {
-    // Search failed — default to ADD
-    return { action: 'new', reason: 'Search unavailable, defaulting to new' };
-  }
-
-  if (hits.length === 0) {
-    return { action: 'new', reason: 'No similar existing memories found' };
-  }
-
-  // LLM-powered resolution (Mem0-style, quality-first)
-  if (useLLM) {
-    const llmResult = await resolveWithLLM(extracted, hits, getObservation);
-    if (llmResult) return llmResult;
-    // LLM failed → fall through to rules-based resolution
-  }
-
-  // Rules-based resolution (free mode fallback)
-  const scored = hits.map(hit => ({
-    hit,
-    ...scoreCandidate(extracted, hit),
-  }));
-
-  // Sort by composite score descending
-  scored.sort((a, b) => b.score - a.score);
-  const best = scored[0];
-
-  // ── Decision logic ──
-
+
+/**
+ * Run Stage 2: Resolve.
+ *
+ * Determines the resolution action for an enriched memory by comparing
+ * it against existing memories found via search.
+ */
+export async function runResolve(
+  extracted: ExtractResult,
+  projectId: string,
+  searchMemories: (query: string, limit: number, projectId: string) => Promise<SearchHit[]>,
+  getObservation: (id: number) => ExistingMemoryRef | null,
+  useLLM = false,
+): Promise<ResolveResult> {
+  // Search for similar existing memories
+  const query = `${extracted.title} ${extracted.narrative.substring(0, 200)}`;
+  let hits: SearchHit[];
+  try {
+    hits = await searchMemories(query, 5, projectId);
+  } catch {
+    // Search failed — default to ADD
+    return { action: 'new', reason: 'Search unavailable, defaulting to new' };
+  }
+
+  if (hits.length === 0) {
+    return { action: 'new', reason: 'No similar existing memories found' };
+  }
+
+  // LLM-powered resolution (Mem0-style, quality-first)
+  if (useLLM) {
+    const llmResult = await resolveWithLLM(extracted, hits, getObservation);
+    if (llmResult) return llmResult;
+    // LLM failed → fall through to rules-based resolution
+  }
+
+  // Rules-based resolution (free mode fallback)
+  const scored = hits.map(hit => ({
+    hit,
+    ...scoreCandidate(extracted, hit),
+  }));
+
+  // Sort by composite score descending
+  scored.sort((a, b) => b.score - a.score);
+  const best = scored[0];
+
+  // ── Decision logic ──
+
   // Very high normalized search similarity → likely duplicate.
   // Raw backend ranking scores must not be compared to 0..1 thresholds.
   if (best.searchSimilarity >= SIMILARITY_DUPLICATE) {
-    if (best.richer) {
-      // New is richer → evolve (supersede)
-      const existing = getObservation(best.hit.observationId);
-      const oldFacts = existing?.facts ?? best.hit.facts.split('\n').filter(Boolean);
-      return {
-        action: 'evolve',
-        targetId: best.hit.observationId,
-        reason: `Near-duplicate of #${best.hit.observationId} but richer content (score: ${best.score.toFixed(2)})`,
-        mergedNarrative: mergeNarratives(best.hit.narrative, extracted.narrative),
-        mergedFacts: mergeFacts(oldFacts, extracted.facts),
-      };
-    }
-    return {
-      action: 'discard',
-      targetId: best.hit.observationId,
-      reason: `Duplicate of #${best.hit.observationId} (score: ${best.score.toFixed(2)})`,
-    };
-  }
-
-  // High similarity → same topic
-  if (best.score >= SIMILARITY_HIGH) {
-    if (best.contradiction) {
-      // Content contradicts existing → evolve
-      const existing = getObservation(best.hit.observationId);
-      const oldFacts = existing?.facts ?? best.hit.facts.split('\n').filter(Boolean);
-      return {
-        action: 'evolve',
-        targetId: best.hit.observationId,
-        reason: `Supersedes #${best.hit.observationId}: contradiction detected (score: ${best.score.toFixed(2)})`,
-        mergedNarrative: extracted.narrative,
-        mergedFacts: mergeFacts(oldFacts, extracted.facts),
-      };
-    }
-
-    if (best.richer) {
-      // New is richer → merge
-      const existing = getObservation(best.hit.observationId);
-      const oldFacts = existing?.facts ?? best.hit.facts.split('\n').filter(Boolean);
-      return {
-        action: 'merge',
-        targetId: best.hit.observationId,
-        reason: `Merging with #${best.hit.observationId}: same topic, new content is richer (score: ${best.score.toFixed(2)})`,
-        mergedNarrative: mergeNarratives(best.hit.narrative, extracted.narrative),
-        mergedFacts: mergeFacts(oldFacts, extracted.facts),
-      };
-    }
-
-    // Not richer → discard
-    return {
-      action: 'discard',
-      targetId: best.hit.observationId,
-      reason: `Already covered by #${best.hit.observationId} (score: ${best.score.toFixed(2)})`,
-    };
-  }
-
-  // Medium similarity + entity match → merge
-  if (best.score >= SIMILARITY_MEDIUM && best.entityMatch) {
-    const existing = getObservation(best.hit.observationId);
-    const oldFacts = existing?.facts ?? best.hit.facts.split('\n').filter(Boolean);
-    const newFactCount = extracted.facts.length;
-    const oldFactCount = oldFacts.length;
-
-    if (newFactCount > oldFactCount) {
-      return {
-        action: 'merge',
-        targetId: best.hit.observationId,
-        reason: `Same entity "${extracted.entityName}", new memory has more facts (${newFactCount} > ${oldFactCount})`,
-        mergedNarrative: mergeNarratives(best.hit.narrative, extracted.narrative),
-        mergedFacts: mergeFacts(oldFacts, extracted.facts),
-      };
-    }
-  }
-
-  // Low similarity or different entity → new memory
-  return { action: 'new', reason: `Different from existing memories (best score: ${best.score.toFixed(2)})` };
-}
+    if (best.richer) {
+      // New is richer → evolve (supersede)
+      const existing = getObservation(best.hit.observationId);
+      const oldFacts = existing?.facts ?? best.hit.facts.split('\n').filter(Boolean);
+      return {
+        action: 'evolve',
+        targetId: best.hit.observationId,
+        reason: `Near-duplicate of #${best.hit.observationId} but richer content (score: ${best.score.toFixed(2)})`,
+        mergedNarrative: mergeNarratives(best.hit.narrative, extracted.narrative),
+        mergedFacts: mergeFacts(oldFacts, extracted.facts),
+      };
+    }
+    return {
+      action: 'discard',
+      targetId: best.hit.observationId,
+      reason: `Duplicate of #${best.hit.observationId} (score: ${best.score.toFixed(2)})`,
+    };
+  }
+
+  // High similarity → same topic
+  if (best.score >= SIMILARITY_HIGH) {
+    if (best.contradiction) {
+      // Content contradicts existing → evolve
+      const existing = getObservation(best.hit.observationId);
+      const oldFacts = existing?.facts ?? best.hit.facts.split('\n').filter(Boolean);
+      return {
+        action: 'evolve',
+        targetId: best.hit.observationId,
+        reason: `Supersedes #${best.hit.observationId}: contradiction detected (score: ${best.score.toFixed(2)})`,
+        mergedNarrative: extracted.narrative,
+        mergedFacts: mergeFacts(oldFacts, extracted.facts),
+      };
+    }
+
+    if (best.richer) {
+      // New is richer → merge
+      const existing = getObservation(best.hit.observationId);
+      const oldFacts = existing?.facts ?? best.hit.facts.split('\n').filter(Boolean);
+      return {
+        action: 'merge',
+        targetId: best.hit.observationId,
+        reason: `Merging with #${best.hit.observationId}: same topic, new content is richer (score: ${best.score.toFixed(2)})`,
+        mergedNarrative: mergeNarratives(best.hit.narrative, extracted.narrative),
+        mergedFacts: mergeFacts(oldFacts, extracted.facts),
+      };
+    }
+
+    // Not richer → discard
+    return {
+      action: 'discard',
+      targetId: best.hit.observationId,
+      reason: `Already covered by #${best.hit.observationId} (score: ${best.score.toFixed(2)})`,
+    };
+  }
+
+  // Medium similarity + entity match → merge
+  if (best.score >= SIMILARITY_MEDIUM && best.entityMatch) {
+    const existing = getObservation(best.hit.observationId);
+    const oldFacts = existing?.facts ?? best.hit.facts.split('\n').filter(Boolean);
+    const newFactCount = extracted.facts.length;
+    const oldFactCount = oldFacts.length;
+
+    if (newFactCount > oldFactCount) {
+      return {
+        action: 'merge',
+        targetId: best.hit.observationId,
+        reason: `Same entity "${extracted.entityName}", new memory has more facts (${newFactCount} > ${oldFactCount})`,
+        mergedNarrative: mergeNarratives(best.hit.narrative, extracted.narrative),
+        mergedFacts: mergeFacts(oldFacts, extracted.facts),
+      };
+    }
+  }
+
+  // Low similarity or different entity → new memory
+  return { action: 'new', reason: `Different from existing memories (best score: ${best.score.toFixed(2)})` };
+}