opcode-pg-memory 2.2.8 → 2.3.0

package/dist/src/mcp/recall-memory.js

@@ -0,0 +1,900 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.recallMemory = recallMemory;
+const logger_1 = require("../services/logger");
+const embedding_1 = require("../utils/embedding");
+const logger = (0, logger_1.createLogger)('recall-memory');
+// ============================================================
+// Constants
+// ============================================================
+const DEFAULT_DECAY_CONFIG = {
+    enabled: true,
+    factor: 0.99,
+    maxAgeDays: 365,
+};
+const DEFAULT_CONFIG = {
+    weights: {
+        semantic: 0.5,
+        recency: 0.3,
+        importance: 0.2,
+    },
+    maxResults: 10,
+    rerankEnabled: true,
+    decay: { ...DEFAULT_DECAY_CONFIG },
+};
+/** Topic-fusion blend ratio: 70% query + 30% topic */
+const TOPIC_FUSION_RATIO = 0.3;
+/** Max results per strategy before merge */
+const PER_STRATEGY_LIMIT = 20;
+// ============================================================
+// Main function: recallMemory
+// ============================================================
+/**
+ * Enhanced recall_memory MCP tool.
+ *
+ * Innovations:
+ * 1. Topic context fusion — blends query embedding with current topic embedding
+ * 2. Structured output — every result carries full context metadata
+ * 3. Agent-friendly — OmO agents pass caller_context for better retrieval
+ * 4. Enhanced filters — min_importance, tier, exclude_topic_segment_ids
+ *
+ * Backward compatible with simple { query: "..." } calls.
+ */
+async function recallMemory(input, pool, config = {}) {
+    const mergedConfig = { ...DEFAULT_CONFIG, ...config };
+    const startTime = Date.now();
+    logger.info(`recall_memory called: "${input.query.substring(0, 100)}..."`);
+    try {
+        // ── Step 0: Resolve session ID ──
+        const sessionId = await resolveSessionId(input, pool);
+        // ── Step 1: Generate query embedding ──
+        const embeddingService = (0, embedding_1.getEmbeddingService)();
+        if (!embeddingService) {
+            throw new Error('Embedding service is not available. Check EMBEDDING_PROVIDER and API keys.');
+        }
+        const queryEmbedding = await embeddingService.generateEmbedding(input.query);
+        // ── Step 2: Topic context fusion ──
+        let fusedEmbedding = queryEmbedding;
+        let contextUsed;
+        const effectiveSessionId = input.caller_context?.current_session_id || input.session_id;
+        if (effectiveSessionId) {
+            const fusionResult = await topicContextFusion(effectiveSessionId, queryEmbedding, embeddingService, pool);
+            if (fusionResult) {
+                fusedEmbedding = fusionResult.fusedEmbedding;
+                contextUsed = fusionResult.contextUsed;
+                logger.info(`Topic fusion applied (topic: ${fusionResult.contextUsed.topic_segment_id.substring(0, 8)}...)`);
+            }
+        }
+        // ── Step 3: Multi-strategy parallel retrieval ──
+        const strategies = input.retrieval_strategies || ['semantic', 'bm25', 'graph'];
+        const retrievalResults = await parallelRetrieve(input.query, fusedEmbedding, sessionId, strategies, pool, input.filters);
+        // ── Step 4: Merge & deduplicate ──
+        const mergedResults = mergeAndDeduplicate(retrievalResults);
+        // ── Step 5: Multi-dimensional scoring (with time-based decay) ──
+        const scoredResults = calculateMultiDimensionalScores(mergedResults, fusedEmbedding, mergedConfig.weights, { ...DEFAULT_DECAY_CONFIG, ...mergedConfig.decay });
+        // ── Step 6: Apply filters ──
+        let filteredResults = applyFilters(scoredResults, input.filters);
+        // ── Step 7: Cross-encoder rerank (backward compat) ──
+        if (input.rerank !== false && mergedConfig.rerankEnabled) {
+            filteredResults = await crossEncoderRerank(filteredResults, input.query);
+        }
+        // ── Step 8: Limit results ──
+        const maxResults = input.max_results || mergedConfig.maxResults;
+        filteredResults = filteredResults.slice(0, maxResults);
+        // ── Step 9: Enrich with context ──
+        const enrichedResults = await enrichWithContext(filteredResults, sessionId, pool);
+        // ── Step 10: Convert to structured MemoryResult ──
+        const results = enrichedResults.map((fact) => {
+            const id = fact.id || `fallback-${Date.now()}-${Math.random().toString(36).slice(2)}`;
+            return {
+                id,
+                type: mapType(fact.type),
+                data: buildDataPayload(fact),
+                relevance_score: fact.relevanceScore,
+                context: {
+                    session_id: fact._sessionId || sessionId,
+                    omo_task_id: fact._omoTaskId,
+                    topic_segment_id: fact._topicSegmentId || 'unknown',
+                    topic_summary: fact._topicSummary,
+                    timestamp: fact._timestamp || new Date().toISOString(),
+                },
+                // backward-compat flat fields
+                content: fact.content,
+                metadata: {
+                    ...fact.metadata,
+                    id,
+                    topic_segment_id: fact._topicSegmentId,
+                    topic_summary: fact._topicSummary,
+                    session_id: fact._sessionId || sessionId,
+                },
+            };
+        });
+        const retrievalTime = Date.now() - startTime;
+        logger.info(`recall_memory completed: ${results.length} results in ${retrievalTime}ms`);
+        return {
+            query: input.query,
+            context_used: contextUsed,
+            success: true,
+            results,
+            total_found: mergedResults.length,
+            retrieval_time_ms: retrievalTime,
+            strategies_used: strategies,
+            session_id: sessionId,
+        };
+    }
+    catch (error) {
+        logger.error('recall_memory error:', error);
+        const retrievalTime = Date.now() - startTime;
+        return {
+            query: input.query,
+            success: false,
+            results: [],
+            total_found: 0,
+            retrieval_time_ms: retrievalTime,
+            strategies_used: [],
+            session_id: '',
+            error: error.message || String(error),
+        };
+    }
+}
+// ============================================================
+// Step 0: Session ID resolution
+// ============================================================
+async function resolveSessionId(input, pool) {
+    if (input.session_id) {
+        // Provided session_id — try session_map first, then fall back to sessions
+        const id = await resolveExternalSessionId(input.session_id, pool);
+        if (id)
+            return id;
+        // If not found in session_map, try legacy sessions table
+        const legacyResult = await pool.query('SELECT id FROM sessions WHERE external_id = $1', [input.session_id]);
+        if (legacyResult.rows.length > 0)
+            return legacyResult.rows[0].id;
+        throw new Error(`Session not found: ${input.session_id}`);
+    }
+    // Auto-detect: try session_map first, then sessions
+    try {
+        const recent = await pool.query('SELECT id, opencode_session_id FROM session_map ORDER BY last_active_at DESC LIMIT 1');
+        if (recent.rows.length > 0) {
+            logger.info(`Auto-detected session: ${recent.rows[0].opencode_session_id}`);
+            return recent.rows[0].id;
+        }
+    }
+    catch {
+        // session_map table doesn't exist yet
+    }
+    // Fall back to legacy sessions
+    const recentSession = await pool.query("SELECT id, external_id FROM sessions ORDER BY updated_at DESC NULLS LAST, created_at DESC LIMIT 1");
+    if (recentSession.rows.length > 0) {
+        logger.info(`Auto-detected session (legacy): ${recentSession.rows[0].external_id}`);
+        return recentSession.rows[0].id;
+    }
+    throw new Error('No session found. Please start a conversation first or provide session_id explicitly.');
+}
+/**
+ * Resolve an external session ID to internal UUID via session_map or sessions table.
+ */
+async function resolveExternalSessionId(externalId, pool) {
+    // Try session_map
+    try {
+        const result = await pool.query('SELECT id FROM session_map WHERE opencode_session_id = $1', [externalId]);
+        if (result.rows.length > 0)
+            return result.rows[0].id;
+    }
+    catch {
+        // table may not exist
+    }
+    // Try legacy sessions
+    const legacy = await pool.query('SELECT id FROM sessions WHERE external_id = $1', [externalId]);
+    if (legacy.rows.length > 0)
+        return legacy.rows[0].id;
+    return null;
+}
+// ============================================================
+// Step 2: Topic context fusion
+// ============================================================
+async function topicContextFusion(externalSessionId, queryEmbedding, embeddingService, pool) {
+    try {
+        // Look up internal session ID
+        const sessionLookup = await pool.query(`SELECT id FROM session_map WHERE opencode_session_id = $1
+       UNION ALL
+       SELECT id FROM sessions WHERE external_id = $1
+       LIMIT 1`, [externalSessionId]);
+        if (sessionLookup.rows.length === 0)
+            return null;
+        const internalSessionId = sessionLookup.rows[0].id;
+        // Find most recent active topic segment for this session
+        const topicResult = await pool.query(`SELECT id, summary, embedding
+       FROM topic_segments
+       WHERE session_map_id = $1
+         AND closed_at IS NULL
+         AND embedding IS NOT NULL
+       ORDER BY created_at DESC
+       LIMIT 1`, [internalSessionId]);
+        if (topicResult.rows.length === 0) {
+            // Try finding any topic segment with embedding
+            const anyTopic = await pool.query(`SELECT id, summary, embedding
+         FROM topic_segments
+         WHERE session_map_id = $1
+           AND embedding IS NOT NULL
+         ORDER BY created_at DESC
+         LIMIT 1`, [internalSessionId]);
+            if (anyTopic.rows.length === 0)
+                return null;
+            const topicEmbedding = anyTopic.rows[0].embedding;
+            const fusedEmbedding = fuseEmbeddings(queryEmbedding, topicEmbedding, TOPIC_FUSION_RATIO);
+            return {
+                fusedEmbedding,
+                contextUsed: {
+                    topic_segment_id: anyTopic.rows[0].id,
+                    topic_summary: anyTopic.rows[0].summary || 'No summary',
+                },
+            };
+        }
+        const topicEmbedding = topicResult.rows[0].embedding;
+        const fusedEmbedding = fuseEmbeddings(queryEmbedding, topicEmbedding, TOPIC_FUSION_RATIO);
+        return {
+            fusedEmbedding,
+            contextUsed: {
+                topic_segment_id: topicResult.rows[0].id,
+                topic_summary: topicResult.rows[0].summary || 'No summary',
+            },
+        };
+    }
+    catch (err) {
+        // topic_segments table may not exist — skip fusion gracefully
+        logger.warn('Topic context fusion skipped:', err);
+        return null;
+    }
+}
+/**
+ * Fuse two embeddings: fused = normalize((1-ratio) * a + ratio * b)
+ */
+function fuseEmbeddings(queryEmb, topicEmb, topicRatio) {
+    const dim = Math.min(queryEmb.length, topicEmb.length);
+    const result = new Array(dim);
+    const queryWeight = 1 - topicRatio;
+    for (let i = 0; i < dim; i++) {
+        result[i] = queryWeight * (queryEmb[i] || 0) + topicRatio * (topicEmb[i] || 0);
+    }
+    return normalizeVector(result);
+}
+/**
+ * L2-normalize a vector to unit length.
+ */
+function normalizeVector(vec) {
+    const magnitude = Math.sqrt(vec.reduce((sum, v) => sum + v * v, 0));
+    if (magnitude === 0)
+        return vec.slice();
+    return vec.map((v) => v / magnitude);
+}
+// ============================================================
+// Step 3: Multi-strategy parallel retrieval
+// ============================================================
+async function parallelRetrieve(query, queryEmbedding, sessionId, strategies, pool, filters) {
+    const results = new Map();
+    const filterSQL = buildFilterConditions(filters);
+    const promises = [];
+    if (strategies.includes('semantic')) {
+        promises.push((async () => {
+            try {
+                const r = await semanticSearch(queryEmbedding, sessionId, pool, filterSQL);
+                results.set('semantic', r);
+            }
+            catch (err) {
+                logger.warn('Semantic search failed:', err);
+                results.set('semantic', []);
+            }
+        })());
+    }
+    if (strategies.includes('bm25')) {
+        promises.push((async () => {
+            try {
+                const r = await bm25Search(query, sessionId, pool, filterSQL);
+                results.set('bm25', r);
+            }
+            catch (err) {
+                logger.warn('BM25 search failed:', err);
+                results.set('bm25', []);
+            }
+        })());
+    }
+    if (strategies.includes('graph')) {
+        promises.push((async () => {
+            try {
+                const r = await graphTraversal(query, embeddingServiceFallback(query), sessionId, pool, filterSQL);
+                results.set('graph', r);
+            }
+            catch (err) {
+                logger.warn('Graph traversal failed:', err);
+                results.set('graph', []);
+            }
+        })());
+    }
+    if (strategies.includes('keyword')) {
+        promises.push((async () => {
+            try {
+                const r = await keywordSearch(query, sessionId, pool, filterSQL);
+                results.set('keyword', r);
+            }
+            catch (err) {
+                logger.warn('Keyword search failed:', err);
+                results.set('keyword', []);
+            }
+        })());
+    }
+    await Promise.all(promises);
+    return results;
+}
+/** Fallback embedding for non-vector strategies (simple keyword affinity) */
+function embeddingServiceFallback(_query) {
+    // Graph traversal uses keyword matching, not vector search
+    return [];
+}
+// ============================================================
+// Filter builder
+// ============================================================
+function buildFilterConditions(filters) {
+    const conditions = [];
+    const params = [];
+    let idx = 1;
+    if (filters?.min_confidence !== undefined) {
+        conditions.push(`(e.confidence >= $${idx} OR o.importance IS NULL)`);
+        params.push(filters.min_confidence);
+        idx++;
+    }
+    // Note: min_importance, tier, entity_types, exclude_topic_segment_ids are applied
+    // in the post-retrieval applyFilters() step, not in SQL.
+    const sql = conditions.length > 0 ? `AND ${conditions.join(' AND ')}` : '';
+    return { sql, params };
+}
+// ============================================================
+// Semantic search (vector)
+// ============================================================
+async function semanticSearch(queryEmbedding, sessionId, pool, filters) {
+    const facts = [];
+    const embeddingStr = formatVectorLiteral(queryEmbedding);
+    // ── 1. Entities (via session_map + topic_segments if available, fall back to sessions) ──
+    try {
+        const entityQuery = `
+      SELECT e.id, e.name, e.type, e.tier, e.weight, e.description,
+             1 - (e.embedding <=> $1) as similarity,
+             e.first_seen_at as created_at, e.confidence,
+             e.session_id, e.session_map_id, e.topic_segment_id
+      FROM entities e
+      LEFT JOIN session_map sm ON e.session_map_id = sm.id
+      LEFT JOIN topic_segments ts ON e.topic_segment_id = ts.id
+      WHERE (e.session_map_id = $2 OR e.session_id = $2 OR e.tier = 'permanent')
+        AND e.embedding IS NOT NULL
+        ${filters.sql}
+      ORDER BY e.embedding <=> $1
+      LIMIT ${PER_STRATEGY_LIMIT}
+    `;
+        const entityResult = await pool.query(entityQuery, [
+            queryEmbedding,
+            sessionId,
+            ...filters.params.filter((_, i) => i < filters.params.length),
+        ]);
+        facts.push(...entityResult.rows.map((row) => ({
+            id: row.id,
+            type: 'entity',
+            content: `[${(row.type || '').toUpperCase()}] ${row.name}: ${row.description || ''}`,
+            relevanceScore: row.similarity ?? 0,
+            tokens: 0,
+            metadata: {
+                entityType: row.type,
+                weight: row.weight,
+                confidence: row.confidence,
+                createdAt: row.created_at,
+                tier: row.tier,
+                source: 'semantic',
+            },
+            _sessionId: row.session_map_id || row.session_id,
+            _topicSegmentId: row.topic_segment_id,
+            _timestamp: row.created_at,
+        })));
+    }
+    catch (err) {
+        logger.warn('Entity semantic search error:', err);
+    }
+    // ── 2. Observations ──
+    try {
+        const obsQuery = `
+      SELECT o.id, o.tool_name, o.tool_output_summary as content, o.embedding,
+             1 - (o.embedding <=> $1) as similarity,
+             o.created_at, o.importance,
+             o.session_id, o.session_map_id, o.topic_segment_id
+      FROM observations o
+      LEFT JOIN session_map sm ON o.session_map_id = sm.id
+      LEFT JOIN topic_segments ts ON o.topic_segment_id = ts.id
+      WHERE (o.session_map_id = $2 OR o.session_id = $2)
+        AND o.embedding IS NOT NULL
+      ORDER BY o.embedding <=> $1
+      LIMIT ${PER_STRATEGY_LIMIT}
+    `;
+        const obsResult = await pool.query(obsQuery, [queryEmbedding, sessionId]);
+        facts.push(...obsResult.rows.map((row) => ({
+            id: row.id,
+            type: 'observation',
+            content: `[${row.tool_name || 'Observation'}] ${row.content || ''}`,
+            relevanceScore: row.similarity ?? 0,
+            tokens: 0,
+            metadata: {
+                toolName: row.tool_name,
+                importance: row.importance,
+                createdAt: row.created_at,
+                source: 'semantic',
+            },
+            _sessionId: row.session_map_id || row.session_id,
+            _topicSegmentId: row.topic_segment_id,
+            _timestamp: row.created_at,
+        })));
+    }
+    catch (err) {
+        logger.warn('Observation semantic search error:', err);
+    }
+    // ── 3. Reflections ──
+    try {
+        const refQuery = `
+      SELECT r.id, r.summary as content, r.pattern_type, r.embedding,
+             1 - (r.embedding <=> $1) as similarity,
+             r.created_at, r.confidence,
+             r.session_id, r.session_map_id, r.topic_segment_id
+      FROM reflections r
+      LEFT JOIN session_map sm ON r.session_map_id = sm.id
+      LEFT JOIN topic_segments ts ON r.topic_segment_id = ts.id
+      WHERE (r.session_map_id = $2 OR r.session_id = $2)
+        AND r.embedding IS NOT NULL
+      ORDER BY r.embedding <=> $1
+      LIMIT ${PER_STRATEGY_LIMIT / 2}
+    `;
+        const refResult = await pool.query(refQuery, [queryEmbedding, sessionId]);
+        facts.push(...refResult.rows.map((row) => ({
+            id: row.id,
+            type: 'reflection',
+            content: `[Reflection${row.pattern_type ? ` - ${row.pattern_type}` : ''}] ${row.content || ''}`,
+            relevanceScore: row.similarity ?? 0,
+            tokens: 0,
+            metadata: {
+                patternType: row.pattern_type,
+                confidence: row.confidence,
+                createdAt: row.created_at,
+                source: 'semantic',
+            },
+            _sessionId: row.session_map_id || row.session_id,
+            _topicSegmentId: row.topic_segment_id,
+            _timestamp: row.created_at,
+        })));
+    }
+    catch (err) {
+        logger.warn('Reflection semantic search error:', err);
+    }
+    return facts;
+}
+// ============================================================
+// BM25 / trigram text search
+// ============================================================
+async function bm25Search(query, sessionId, pool, filters) {
+    const queryTerms = query.split(/\s+/).filter((t) => t.length > 2);
+    if (queryTerms.length === 0)
+        return [];
+    const facts = [];
+    try {
+        const entityQuery = `
+      SELECT e.id, e.name, e.type, e.tier, e.weight, e.description,
+             similarity(e.name, $1) as bm25_score,
+             e.first_seen_at as created_at, e.confidence,
+             e.session_id, e.session_map_id, e.topic_segment_id
+      FROM entities e
+      WHERE (e.session_map_id = $2 OR e.session_id = $2 OR e.tier = 'permanent')
+        AND (e.name % $1 OR e.description % $1)
+        ${filters.sql}
+      ORDER BY similarity(e.name, $1) DESC
+      LIMIT ${PER_STRATEGY_LIMIT}
+    `;
+        const entityResult = await pool.query(entityQuery, [
+            query,
+            sessionId,
+            ...filters.params,
+        ]);
+        facts.push(...entityResult.rows.map((row) => ({
+            id: row.id,
+            type: 'entity',
+            content: `[${(row.type || '').toUpperCase()}] ${row.name}: ${row.description || ''}`,
+            relevanceScore: row.bm25_score ?? 0,
+            tokens: 0,
+            metadata: {
+                entityType: row.type,
+                weight: row.weight,
+                confidence: row.confidence,
+                createdAt: row.created_at,
+                tier: row.tier,
+                source: 'bm25',
+            },
+            _sessionId: row.session_map_id || row.session_id,
+            _topicSegmentId: row.topic_segment_id,
+            _timestamp: row.created_at,
+        })));
+    }
+    catch (err) {
+        logger.warn('BM25 search failed (pg_trgm may not be installed):', err);
+    }
+    return facts;
+}
+// ============================================================
+// Graph traversal
+// ============================================================
+async function graphTraversal(query, _queryEmb, sessionId, pool, filters) {
+    // 1. Find seed entities matching query text
+    let seedQuery;
+    let seedParams;
+    try {
+        // Try newer schema (session_map_id)
+        seedQuery = `
+      SELECT id, name, type
+      FROM entities
+      WHERE (session_map_id = $1 OR session_id = $1 OR tier = 'permanent')
+        AND (name ILIKE $2 OR description ILIKE $2)
+      LIMIT 5
+    `;
+        seedParams = [sessionId, `%${query}%`];
+    }
+    catch {
+        return [];
+    }
+    const seedResult = await pool.query(seedQuery, seedParams);
+    if (seedResult.rows.length === 0)
+        return [];
+    const seedIds = seedResult.rows.map((row) => row.id);
+    // 2. Traverse relations (1-hop neighbors) using the graph query pattern from spec
+    const graphQuery = `
+    SELECT
+      e2.id, e2.name, e2.type, e2.tier, e2.weight, e2.description,
+      e2.first_seen_at as created_at, e2.confidence,
+      r.relation_type,
+      e_seed.name as related_entity_name,
+      e2.session_id, e2.session_map_id, e2.topic_segment_id
+    FROM entities e_seed
+    JOIN relations r ON e_seed.id = r.source_entity_id
+    JOIN entities e2 ON r.target_entity_id = e2.id
+    WHERE e_seed.id = ANY($1)
+      AND r.confidence >= $2
+      AND e_seed.id != e2.id
+    ORDER BY r.confidence DESC
+    LIMIT ${PER_STRATEGY_LIMIT}
+  `;
+    try {
+        const graphResult = await pool.query(graphQuery, [seedIds, 0.5]);
+        return graphResult.rows.map((row) => ({
+            id: row.id,
+            type: 'entity',
+            content: `[${(row.type || '').toUpperCase()}] ${row.name}: ${row.description || ''} (related to ${row.related_entity_name} via ${row.relation_type})`,
+            relevanceScore: (row.confidence ?? 0) * 0.8,
+            tokens: 0,
+            metadata: {
+                entityType: row.type,
+                weight: row.weight,
+                confidence: row.confidence,
+                createdAt: row.created_at,
+                tier: row.tier,
+                relationType: row.relation_type,
+                relatedEntity: row.related_entity_name,
+                source: 'graph',
+            },
+            _sessionId: row.session_map_id || row.session_id,
+            _topicSegmentId: row.topic_segment_id,
+            _timestamp: row.created_at,
+        }));
+    }
+    catch (err) {
+        logger.warn('Graph traversal error:', err);
+        return [];
+    }
+}
+// ============================================================
+// Keyword search
+// ============================================================
+async function keywordSearch(query, sessionId, pool, filters) {
+    const keywords = query.split(/\s+/).filter((k) => k.length > 2);
+    if (keywords.length === 0)
+        return [];
+    const pattern = keywords.join('|');
+    try {
+        const entityQuery = `
+      SELECT e.id, e.name, e.type, e.tier, e.weight, e.description,
+             e.first_seen_at as created_at, e.confidence,
+             e.session_id, e.session_map_id, e.topic_segment_id
+      FROM entities e
+      WHERE (e.session_map_id = $1 OR e.session_id = $1 OR e.tier = 'permanent')
+        AND (e.name ~* $2 OR e.description ~* $2)
+        ${filters.sql}
+      LIMIT ${PER_STRATEGY_LIMIT}
+    `;
+        const entityResult = await pool.query(entityQuery, [sessionId, pattern, ...filters.params]);
+        return entityResult.rows.map((row) => ({
+            id: row.id,
+            type: 'entity',
+            content: `[${(row.type || '').toUpperCase()}] ${row.name}: ${row.description || ''}`,
+            relevanceScore: 0.6,
+            tokens: 0,
+            metadata: {
+                entityType: row.type,
+                weight: row.weight,
+                confidence: row.confidence,
+                createdAt: row.created_at,
+                tier: row.tier,
+                source: 'keyword',
+            },
+            _sessionId: row.session_map_id || row.session_id,
+            _topicSegmentId: row.topic_segment_id,
+            _timestamp: row.created_at,
+        }));
+    }
+    catch (err) {
+        logger.warn('Keyword search error:', err);
+        return [];
+    }
+}
+// ============================================================
+// Step 4: Merge & deduplicate
+// ============================================================
+function mergeAndDeduplicate(strategyResults) {
+    const seen = new Set();
+    const merged = [];
+    for (const [, facts] of strategyResults) {
+        for (const fact of facts) {
+            const key = `${fact.type}:${fact.id || fact.metadata.id || ''}`;
+            if (!seen.has(key)) {
+                seen.add(key);
+                merged.push(fact);
+            }
+            else {
+                // Keep highest score
+                const existing = merged.find((f) => `${f.type}:${f.id || f.metadata.id || ''}` === key);
+                if (existing && fact.relevanceScore > existing.relevanceScore) {
+                    existing.relevanceScore = fact.relevanceScore;
+                }
+            }
+        }
+    }
+    return merged;
+}
+// ============================================================
+// Step 5: Multi-dimensional scoring
+// ============================================================
+function calculateMultiDimensionalScores(facts, _queryEmbedding, weights, decayConfig = DEFAULT_DECAY_CONFIG) {
+    const now = new Date();
+    let results = facts
+        .map((fact) => {
+        const semanticScore = fact.relevanceScore;
+        // Recency decay
+        const createdAt = new Date(fact.metadata.createdAt || fact._timestamp || Date.now());
+        const daysAgo = (now.getTime() - createdAt.getTime()) / (1000 * 60 * 60 * 24);
+        const recencyScore = 1.0 / (1 + daysAgo);
+        // Importance (normalize to 0-1)
+        const importance = fact.metadata.importance || fact.metadata.weight || 3;
+        const importanceScore = Math.min(1, importance / 5.0);
+        const finalScore = weights.semantic * semanticScore +
+            weights.recency * recencyScore +
+            weights.importance * importanceScore;
+        let relevanceScore = Math.round(finalScore * 1000) / 1000;
+        // ── Apply time-based entity weight decay ──
+        if (decayConfig.enabled) {
+            const metadata = fact.metadata || {};
+            const lastSeen = metadata.createdAt || fact._timestamp || Date.now();
+            const daysSinceLastSeen = (Date.now() - new Date(lastSeen).getTime()) / 86400000;
+            // Skip decay for permanent tier
+            if (metadata.tier !== 'permanent') {
+                // Apply exponential decay to relevanceScore
+                const decayedWeight = relevanceScore * Math.pow(decayConfig.factor, daysSinceLastSeen);
+                relevanceScore = Math.max(0.01, decayedWeight);
+                // Mark old, low-value results for filtering
+                if (daysSinceLastSeen > decayConfig.maxAgeDays && relevanceScore < 0.1) {
+                    relevanceScore = 0;
+                }
+            }
+        }
+        return {
+            ...fact,
+            relevanceScore,
+        };
+    })
+        .filter((r) => r.relevanceScore > 0)
+        .sort((a, b) => b.relevanceScore - a.relevanceScore);
+    return results;
+}
+// ============================================================
+// Step 6: Apply filters (post-retrieval)
+// ============================================================
+function applyFilters(facts, filters) {
+    if (!filters)
+        return facts;
+    // Resolve tier filter: single `tier` takes precedence, fall back to `tier_levels` (first element)
+    const effectiveTiers = [];
+    if (filters.tier) {
+        effectiveTiers.push(filters.tier);
+    }
+    else if (filters.tier_levels && filters.tier_levels.length > 0) {
+        effectiveTiers.push(...filters.tier_levels);
+    }
+    const now = Date.now();
+    return facts.filter((fact) => {
+        // min_confidence
+        if (filters.min_confidence !== undefined) {
+            const conf = fact.metadata.confidence;
+            if (conf !== undefined && conf !== null && conf < filters.min_confidence)
+                return false;
+        }
+        // min_importance
+        if (filters.min_importance !== undefined) {
+            const imp = fact.metadata.importance;
+            if (imp !== undefined && imp !== null && imp < filters.min_importance)
+                return false;
+        }
+        // tier filter (single + array backward compat)
+        if (effectiveTiers.length > 0) {
+            const tier = fact.metadata.tier;
+            if (tier && !effectiveTiers.includes(tier))
+                return false;
+        }
+        // entity_types filter
+        if (filters.entity_types && filters.entity_types.length > 0) {
+            const etype = fact.metadata.entityType;
+            if (etype && !filters.entity_types.includes(etype))
+                return false;
+        }
+        // exclude_topic_segment_ids
+        if (filters.exclude_topic_segment_ids && filters.exclude_topic_segment_ids.length > 0) {
+            if (fact._topicSegmentId &&
+                filters.exclude_topic_segment_ids.includes(fact._topicSegmentId)) {
+                return false;
+            }
+        }
+        // time_range_days (backward compat)
+        if (filters.time_range_days !== undefined) {
+            const ts = fact._timestamp || fact.metadata.createdAt;
+            if (ts) {
+                const ageMs = now - new Date(ts).getTime();
+                const ageDays = ageMs / (1000 * 60 * 60 * 24);
+                if (ageDays > filters.time_range_days)
+                    return false;
+            }
+        }
+        return true;
+    });
+}
+// ============================================================
+// Step 7: Cross-encoder rerank (simplified)
+// ============================================================
+async function crossEncoderRerank(facts, query) {
+    const queryLower = query.toLowerCase();
+    const reranked = facts.map((fact) => {
+        const contentLower = fact.content.toLowerCase();
+        const queryWords = queryLower.split(/\s+/);
+        const contentWords = contentLower.split(/\s+/);
+        let overlap = 0;
+        for (const word of queryWords) {
+            if (word.length > 2 && contentWords.some((cw) => cw.includes(word))) {
+                overlap++;
+            }
+        }
+        const overlapScore = queryWords.length > 0 ? overlap / queryWords.length : 0;
+        const adjustedScore = fact.relevanceScore * 0.7 + overlapScore * 0.3;
+        return {
+            ...fact,
+            relevanceScore: Math.round(adjustedScore * 1000) / 1000,
+        };
+    });
+    return reranked.sort((a, b) => b.relevanceScore - a.relevanceScore);
+}
+// ============================================================
+// Step 9: Enrich with context (session_map + topic_segments)
+// ============================================================
+async function enrichWithContext(facts, sessionId, pool) {
+    if (facts.length === 0)
+        return facts;
+    // Collect unique session IDs that need context lookup
+    const sessionIds = new Set();
+    const topicIds = new Set();
+    for (const fact of facts) {
+        if (fact._sessionId)
+            sessionIds.add(fact._sessionId);
+        if (fact._topicSegmentId)
+            topicIds.add(fact._topicSegmentId);
+    }
+    // Batch lookups
+    const [sessionMap, topicMap] = await Promise.all([
+        lookupSessions([...sessionIds], pool),
+        lookupTopics([...topicIds], pool),
+    ]);
+    // Apply context to each fact
+    for (const fact of facts) {
+        if (fact._sessionId && sessionMap.has(fact._sessionId)) {
+            const s = sessionMap.get(fact._sessionId);
+            fact._omoTaskId = fact._omoTaskId || s.omo_task_id;
+            if (!fact._sessionId)
+                fact._sessionId = s.opencode_session_id || fact._sessionId;
+        }
+        if (fact._topicSegmentId && topicMap.has(fact._topicSegmentId)) {
+            const t = topicMap.get(fact._topicSegmentId);
+            fact._topicSummary = t.summary;
+        }
+        // Fallback timestamp
+        if (!fact._timestamp) {
+            fact._timestamp = fact.metadata.createdAt || new Date().toISOString();
+        }
+    }
+    return facts;
+}
+async function lookupSessions(ids, pool) {
+    const map = new Map();
+    if (ids.length === 0)
+        return map;
+    try {
+        const result = await pool.query(`SELECT id, opencode_session_id, omo_task_id FROM session_map WHERE id = ANY($1)`, [ids]);
+        for (const row of result.rows) {
+            map.set(row.id, {
+                opencode_session_id: row.opencode_session_id,
+                omo_task_id: row.omo_task_id,
+            });
+        }
+    }
+    catch {
+        // session_map may not exist, try sessions
+        try {
+            const result = await pool.query(`SELECT id, external_id FROM sessions WHERE id = ANY($1)`, [ids]);
+            for (const row of result.rows) {
+                map.set(row.id, { opencode_session_id: row.external_id });
+            }
+        }
+        catch {
+            // Both failed — context will be sparse
+        }
+    }
+    return map;
+}
+async function lookupTopics(ids, pool) {
+    const map = new Map();
+    if (ids.length === 0)
+        return map;
+    try {
+        const result = await pool.query(`SELECT id, summary FROM topic_segments WHERE id = ANY($1)`, [ids]);
+        for (const row of result.rows) {
+            map.set(row.id, { summary: row.summary });
+        }
+    }
+    catch {
+        // topic_segments may not exist
+    }
+    return map;
+}
+// ============================================================
+// Helpers: type mapping, data payload builder, vector formatting
+// ============================================================
+function mapType(type) {
+    if (type === 'message')
+        return 'observation'; // legacy messages → observation
+    if (type === 'relation')
+        return 'relation';
+    return type;
+}
+function buildDataPayload(fact) {
+    return {
+        name: fact.metadata.entityType,
+        type: fact.type,
+        content: fact.content,
+        confidence: fact.metadata.confidence,
+        importance: fact.metadata.importance || fact.metadata.weight,
+        created_at: fact.metadata.createdAt || fact._timestamp,
+        pattern_type: fact.metadata.patternType,
+        relation_type: fact.metadata.relationType,
+        related_entity: fact.metadata.relatedEntity,
+        tool_name: fact.metadata.toolName,
+        ...fact.metadata,
+    };
+}
+/** Format a number[] as a pgvector literal string: '[0.1,0.2,...]' */
+function formatVectorLiteral(embedding) {
+    return `[${embedding.join(',')}]`;
+}
+//# sourceMappingURL=recall-memory.js.map