@exaudeus/memory-mcp 1.6.0 → 1.7.0

package/dist/index.js

@@ -909,6 +909,7 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
                 // --- Search mode: context provided → keyword search across all topics ---
                 const max = maxResults ?? 10;
                 const threshold = minMatch ?? 0.2;
+                // Resolve which lobes to search — follows the degradation ladder via resolveLobesForRead().
                 const allLobeResults = [];
                 const ctxEntryLobeMap = new Map(); // entry id → lobe name
                 let label;

package/dist/ranking.d.ts

@@ -0,0 +1,57 @@
+import type { MemoryEntry, EmbeddingVector } from './types.js';
+/** Which ranking signal produced this result. */
+export type RankSource = 'keyword' | 'semantic' | 'merged';
+/** A scored search result. Shared return type for all ranking functions. */
+export interface ScoredEntry {
+    readonly entry: MemoryEntry;
+    readonly score: number;
+    readonly matchedKeywords: readonly string[];
+    readonly source: RankSource;
+    /** Raw cosine similarity before boost multiplication.
+     *  Present for semantic and merged results; absent for keyword-only.
+     *  Used for debug logging (threshold calibration) and display. */
+    readonly semanticSimilarity?: number;
+}
+/** Shared context for ranking functions — pure data, no callbacks.
+ *  Groups the parameters that keywordRank and semanticRank need,
+ *  keeping function signatures tight.
+ *
+ *  freshEntryIds is precomputed by the store — the ranking function checks
+ *  set membership rather than calling into the store's private staleness logic.
+ *  This keeps the function provably pure.
+ *
+ *  defaultModuleBoost is the fallback boost for modules/* topics not in topicBoost.
+ *  Injected here so ranking.ts needs zero direct threshold imports. */
+export interface RankContext {
+    readonly currentBranch: string;
+    readonly branchFilter: string | undefined;
+    readonly topicBoost: Readonly<Record<string, number>>;
+    readonly freshEntryIds: ReadonlySet<string>;
+    readonly defaultModuleBoost: number;
+}
+/** Rank entries by keyword overlap with context keywords.
+ *  Pure extraction of the ranking logic from contextSearch — identical scoring.
+ *
+ *  Filter + rank in one pass for efficiency (~200 entries, not worth two iterations).
+ *  Branch filtering for recent-work is applied here because it's a pre-condition
+ *  for ranking, not a separate pipeline stage.
+ *
+ *  Does NOT include the "always include user entries" policy — that's an
+ *  orchestration concern that stays in contextSearch. */
+export declare function keywordRank(entries: readonly MemoryEntry[], contextKeywords: ReadonlySet<string>, minMatch: number, ctx: RankContext): readonly ScoredEntry[];
+/** Rank entries by cosine similarity between query embedding and stored vectors.
+ *  Pure function — no I/O, no side effects.
+ *
+ *  Entries without vectors are silently skipped — they participate via keyword ranking only.
+ *  Branch filtering applied (recent-work scoped to current branch).
+ *
+ *  @param minSimilarity Minimum cosine similarity to include. Caller provides the
+ *    threshold (SEMANTIC_MIN_SIMILARITY for production, 0 for debug mode to see all scores). */
+export declare function semanticRank(entries: readonly MemoryEntry[], vectors: ReadonlyMap<string, EmbeddingVector>, queryVector: EmbeddingVector, minSimilarity: number, ctx: RankContext): readonly ScoredEntry[];
+/** Merge keyword and semantic ranking results using max-score strategy.
+ *  For each entry: final score = max(keywordScore, semanticScore).
+ *  Entries in both lists get source: 'merged', preserving matchedKeywords from
+ *  keyword result and semanticSimilarity from semantic result.
+ *
+ *  No weighted fusion, no magic constants. Deterministic — same inputs, same output. */
+export declare function mergeRankings(keywordResults: readonly ScoredEntry[], semanticResults: readonly ScoredEntry[]): readonly ScoredEntry[];

package/dist/ranking.js

@@ -0,0 +1,140 @@
+// Domain ranking functions — scoring MemoryEntries using text analysis primitives.
+// Pure functions: no I/O, no side effects, deterministic.
+//
+// Separated from text-analyzer.ts (which works on strings/sets, not domain types)
+// and from store.ts (which handles orchestration and persistence).
+// This module is the ranking pipeline seam for keyword, semantic, and merged ranking.
+import { REFERENCE_BOOST_MULTIPLIER, TAG_MATCH_BOOST, } from './thresholds.js';
+import { extractKeywords, stem, cosineSimilarity } from './text-analyzer.js';
+// ─── Shared helpers ────────────────────────────────────────────────────────
+/** Resolve topic boost for an entry. Falls back to defaultModuleBoost for
+ *  modules/* topics, or 1.0 for unknown topics. */
+function getTopicBoost(topic, ctx) {
+    return ctx.topicBoost[topic] ?? (topic.startsWith('modules/') ? ctx.defaultModuleBoost : 1.0);
+}
+/** Check whether a recent-work entry should be filtered out by branch. */
+function isBranchFiltered(entry, ctx) {
+    return entry.topic === 'recent-work'
+        && ctx.branchFilter !== '*'
+        && !!entry.branch
+        && entry.branch !== ctx.currentBranch;
+}
+// ─── Keyword ranking ───────────────────────────────────────────────────────
+/** Rank entries by keyword overlap with context keywords.
+ *  Pure extraction of the ranking logic from contextSearch — identical scoring.
+ *
+ *  Filter + rank in one pass for efficiency (~200 entries, not worth two iterations).
+ *  Branch filtering for recent-work is applied here because it's a pre-condition
+ *  for ranking, not a separate pipeline stage.
+ *
+ *  Does NOT include the "always include user entries" policy — that's an
+ *  orchestration concern that stays in contextSearch. */
+export function keywordRank(entries, contextKeywords, minMatch, ctx) {
+    const results = [];
+    for (const entry of entries) {
+        if (isBranchFiltered(entry, ctx))
+            continue;
+        // Include tag values as keywords so tagged entries surface in context search
+        const tagKeywordPart = entry.tags ? ` ${entry.tags.join(' ')}` : '';
+        const entryKeywords = extractKeywords(`${entry.title} ${entry.content}${tagKeywordPart}`);
+        const matchedKeywords = [];
+        for (const kw of contextKeywords) {
+            if (entryKeywords.has(kw))
+                matchedKeywords.push(kw);
+        }
+        if (matchedKeywords.length === 0)
+            continue;
+        // Enforce minimum match threshold
+        const matchRatio = matchedKeywords.length / contextKeywords.size;
+        if (matchRatio < minMatch)
+            continue;
+        // Score = keyword match ratio × confidence × topic boost × freshness × reference boost × tag boost
+        const boost = getTopicBoost(entry.topic, ctx);
+        const freshnessMultiplier = ctx.freshEntryIds.has(entry.id) ? 1.0 : 0.7;
+        // Reference boost: exact class/file name match in references gets a 1.3x multiplier
+        const referenceBoost = entry.references?.some(ref => {
+            const basename = ref.split('/').pop()?.replace(/\.\w+$/, '') ?? ref;
+            return contextKeywords.has(stem(basename.toLowerCase()));
+        }) ? REFERENCE_BOOST_MULTIPLIER : 1.0;
+        // Tag boost: if any tag exactly matches a context keyword, boost the entry
+        const tagBoost = entry.tags?.some(tag => contextKeywords.has(tag))
+            ? TAG_MATCH_BOOST : 1.0;
+        const score = matchRatio * entry.confidence * boost * freshnessMultiplier * referenceBoost * tagBoost;
+        results.push({ entry, score, matchedKeywords, source: 'keyword' });
+    }
+    return results.sort((a, b) => b.score - a.score);
+}
+// ─── Semantic ranking ──────────────────────────────────────────────────────
+/** Rank entries by cosine similarity between query embedding and stored vectors.
+ *  Pure function — no I/O, no side effects.
+ *
+ *  Entries without vectors are silently skipped — they participate via keyword ranking only.
+ *  Branch filtering applied (recent-work scoped to current branch).
+ *
+ *  @param minSimilarity Minimum cosine similarity to include. Caller provides the
+ *    threshold (SEMANTIC_MIN_SIMILARITY for production, 0 for debug mode to see all scores). */
+export function semanticRank(entries, vectors, queryVector, minSimilarity, ctx) {
+    const results = [];
+    for (const entry of entries) {
+        if (isBranchFiltered(entry, ctx))
+            continue;
+        const entryVector = vectors.get(entry.id);
+        if (!entryVector)
+            continue;
+        const similarity = cosineSimilarity(queryVector, entryVector);
+        if (similarity < minSimilarity)
+            continue;
+        // Score = cosine similarity × confidence × topic boost × freshness
+        // No reference/tag boost — those are keyword-domain signals captured by keywordRank
+        const boost = getTopicBoost(entry.topic, ctx);
+        const freshnessMultiplier = ctx.freshEntryIds.has(entry.id) ? 1.0 : 0.7;
+        const score = similarity * entry.confidence * boost * freshnessMultiplier;
+        results.push({
+            entry,
+            score,
+            matchedKeywords: [],
+            source: 'semantic',
+            semanticSimilarity: similarity,
+        });
+    }
+    return results.sort((a, b) => b.score - a.score);
+}
+// ─── Merge ─────────────────────────────────────────────────────────────────
+/** Merge keyword and semantic ranking results using max-score strategy.
+ *  For each entry: final score = max(keywordScore, semanticScore).
+ *  Entries in both lists get source: 'merged', preserving matchedKeywords from
+ *  keyword result and semanticSimilarity from semantic result.
+ *
+ *  No weighted fusion, no magic constants. Deterministic — same inputs, same output. */
+export function mergeRankings(keywordResults, semanticResults) {
+    // Index keyword results by entry ID for O(1) lookup during merge
+    const keywordById = new Map();
+    for (const r of keywordResults) {
+        keywordById.set(r.entry.id, r);
+    }
+    const merged = new Map();
+    // Process semantic results — check for keyword counterpart
+    for (const sem of semanticResults) {
+        const kw = keywordById.get(sem.entry.id);
+        if (kw) {
+            // Entry in both lists — use max score, merge signals
+            merged.set(sem.entry.id, {
+                entry: sem.entry,
+                score: Math.max(sem.score, kw.score),
+                matchedKeywords: kw.matchedKeywords, // from keyword (semantic has none)
+                source: 'merged',
+                semanticSimilarity: sem.semanticSimilarity, // from semantic
+            });
+            keywordById.delete(sem.entry.id); // consumed
+        }
+        else {
+            // Semantic-only
+            merged.set(sem.entry.id, sem);
+        }
+    }
+    // Remaining keyword-only results
+    for (const kw of keywordById.values()) {
+        merged.set(kw.entry.id, kw);
+    }
+    return Array.from(merged.values()).sort((a, b) => b.score - a.score);
+}

package/dist/store.d.ts

@@ -1,4 +1,5 @@
 import type { MemoryEntry, TopicScope, TrustLevel, DetailLevel, DurabilityDecision, QueryResult, StoreResult, CorrectResult, MemoryStats, BriefingResult, ConflictPair, MemoryConfig } from './types.js';
+import type { ScoredEntry } from './ranking.js';
 export declare class MarkdownMemoryStore {
     private readonly config;
     private readonly memoryPath;
@@ -29,13 +30,14 @@ export declare class MarkdownMemoryStore {
     stats(): Promise<MemoryStats>;
     /** Bootstrap: scan repo structure and seed initial knowledge */
     bootstrap(): Promise<StoreResult[]>;
-    /** Search across all topics using keyword matching with topic-based boosting.
+    /** Search across all topics using keyword + semantic ranking with topic-based boosting.
+     *  Orchestrator: reload, extract keywords, embed query, rank, merge, apply policies.
+     *
+     *  Graceful degradation: when embedder is null or embed fails, semantic results
+     *  are empty and merge produces keyword-only results — identical to pre-embedding behavior.
+     *
      *  @param minMatch Minimum ratio of context keywords that must match (0-1, default 0.2) */
-    contextSearch(context: string, maxResults?: number, branchFilter?: string, minMatch?: number): Promise<Array<{
-        entry: MemoryEntry;
-        score: number;
-        matchedKeywords: string[];
-    }>>;
+    contextSearch(context: string, maxResults?: number, branchFilter?: string, minMatch?: number): Promise<readonly ScoredEntry[]>;
     /** Generate a collision-resistant ID: {prefix}-{8 random hex chars} */
     private generateId;
     /** Compute relative file path for an entry within the memory directory */
@@ -93,6 +95,13 @@ export declare class MarkdownMemoryStore {
     /** Find entries in the same topic with significant overlap (dedup detection).
      *  Uses hybrid jaccard+containment similarity. */
     private findRelatedEntries;
+    /** Find semantic duplicates by cosine similarity against stored vectors.
+     *  Same-topic only. Returns entries above DEDUP_SEMANTIC_THRESHOLD.
+     *  Returns empty when no embedder or no vectors available. */
+    private findSemanticDuplicates;
+    /** Merge keyword-based and semantic-based dedup results, dedup by ID,
+     *  keeping the entry with higher similarity score. */
+    private mergeRelatedEntries;
     /** Tag frequency across all entries — for vocabulary echo in store responses.
      *  Returns tags sorted by frequency (descending). O(N) over entries. */
     getTagFrequency(): ReadonlyMap<string, number>;

package/dist/store.js

@@ -7,9 +7,10 @@ import crypto from 'crypto';
 import { execFile } from 'child_process';
 import { promisify } from 'util';
 import { DEFAULT_CONFIDENCE, realClock, parseTopicScope, parseTrustLevel, parseTags, asEmbeddingVector } from './types.js';
-import { DEDUP_SIMILARITY_THRESHOLD, CONFLICT_SIMILARITY_THRESHOLD_SAME_TOPIC, CONFLICT_SIMILARITY_THRESHOLD_CROSS_TOPIC, CONFLICT_MIN_CONTENT_CHARS, OPPOSITION_PAIRS, PREFERENCE_SURFACE_THRESHOLD, REFERENCE_BOOST_MULTIPLIER, TOPIC_BOOST, MODULE_TOPIC_BOOST, USER_ALWAYS_INCLUDE_SCORE_FRACTION, DEFAULT_STALE_DAYS_STANDARD, DEFAULT_STALE_DAYS_PREFERENCES, DEFAULT_MAX_STALE_IN_BRIEFING, DEFAULT_MAX_DEDUP_SUGGESTIONS, DEFAULT_MAX_CONFLICT_PAIRS, DEFAULT_MAX_PREFERENCE_SUGGESTIONS, TAG_MATCH_BOOST, } from './thresholds.js';
+import { DEDUP_SIMILARITY_THRESHOLD, DEDUP_SEMANTIC_THRESHOLD, SEMANTIC_MIN_SIMILARITY, CONFLICT_SIMILARITY_THRESHOLD_SAME_TOPIC, CONFLICT_SIMILARITY_THRESHOLD_CROSS_TOPIC, CONFLICT_MIN_CONTENT_CHARS, OPPOSITION_PAIRS, PREFERENCE_SURFACE_THRESHOLD, TOPIC_BOOST, MODULE_TOPIC_BOOST, USER_ALWAYS_INCLUDE_SCORE_FRACTION, DEFAULT_STALE_DAYS_STANDARD, DEFAULT_STALE_DAYS_PREFERENCES, DEFAULT_MAX_STALE_IN_BRIEFING, DEFAULT_MAX_DEDUP_SUGGESTIONS, DEFAULT_MAX_CONFLICT_PAIRS, DEFAULT_MAX_PREFERENCE_SUGGESTIONS, } from './thresholds.js';
 import { realGitService } from './git-service.js';
-import { extractKeywords, stem, similarity, matchesFilter, computeRelevanceScore, } from './text-analyzer.js';
+import { extractKeywords, similarity, cosineSimilarity, matchesFilter, computeRelevanceScore, } from './text-analyzer.js';
+import { keywordRank, semanticRank, mergeRankings } from './ranking.js';
 import { detectEphemeralSignals, formatEphemeralWarning, getEphemeralSeverity } from './ephemeral.js';
 // Used only by bootstrap() for git log — not part of the GitService boundary
 // because bootstrap is a one-shot utility, not a recurring operation
@@ -101,8 +102,22 @@ export class MarkdownMemoryStore {
         this.entries.set(id, entry);
         const file = this.entryToRelativePath(entry);
         await this.persistEntry(entry);
-        // Dedup: find related entries in the same topic (excluding the one just stored and any overwritten)
-        const relatedEntries = this.findRelatedEntries(entry, existing?.id);
+        // Embed and persist vector — awaited so .vec exists when store() returns
+        if (this.embedder) {
+            const embedText = `${title}\n\n${content}`;
+            const embedResult = await this.embedder.embed(embedText);
+            if (embedResult.ok) {
+                await this.persistVector(file, embedResult.vector);
+                this.vectors.set(entry.id, embedResult.vector);
+            }
+            else {
+                process.stderr.write(`[memory-mcp] Embedding failed for ${entry.id}: ${embedResult.failure.kind}\n`);
+            }
+        }
+        // Dedup: merge keyword-based and semantic-based duplicate detection
+        const keywordDupes = this.findRelatedEntries(entry, existing?.id);
+        const semanticDupes = this.findSemanticDuplicates(entry.id, topic);
+        const relatedEntries = this.mergeRelatedEntries(keywordDupes, semanticDupes);
         // Surface relevant preferences if storing a non-preference entry
         const relevantPreferences = (topic !== 'preferences' && topic !== 'user')
             ? this.findRelevantPreferences(entry)
@@ -306,6 +321,19 @@ export class MarkdownMemoryStore {
         };
         this.entries.set(id, updated);
         await this.persistEntry(updated);
+        // Re-embed: content changed, old vector is stale
+        if (this.embedder) {
+            const embedText = `${updated.title}\n\n${updated.content}`;
+            const embedResult = await this.embedder.embed(embedText);
+            if (embedResult.ok) {
+                const updatedFile = this.entryToRelativePath(updated);
+                await this.persistVector(updatedFile, embedResult.vector);
+                this.vectors.set(updated.id, embedResult.vector);
+            }
+            else {
+                process.stderr.write(`[memory-mcp] Re-embedding failed for ${updated.id}: ${embedResult.failure.kind}\n`);
+            }
+        }
         return { corrected: true, id, action, newConfidence: 1.0, trust: 'user' };
     }
     /** Get memory health statistics */
@@ -417,59 +445,77 @@ export class MarkdownMemoryStore {
         return results;
     }
     // --- Contextual search (memory_context) ---
-    /** Search across all topics using keyword matching with topic-based boosting.
+    /** Search across all topics using keyword + semantic ranking with topic-based boosting.
+     *  Orchestrator: reload, extract keywords, embed query, rank, merge, apply policies.
+     *
+     *  Graceful degradation: when embedder is null or embed fails, semantic results
+     *  are empty and merge produces keyword-only results — identical to pre-embedding behavior.
+     *
      *  @param minMatch Minimum ratio of context keywords that must match (0-1, default 0.2) */
     async contextSearch(context, maxResults = 10, branchFilter, minMatch = 0.2) {
         // Reload from disk to pick up changes from other processes
         await this.reloadFromDisk();
         const contextKeywords = extractKeywords(context);
-        if (contextKeywords.size === 0)
+        // Only bail on zero keywords when there's no embedder to fall back on.
+        // Stopword-heavy queries produce zero keywords but can yield semantic results.
+        if (contextKeywords.size === 0 && !this.embedder)
             return [];
         const currentBranch = branchFilter || await this.getCurrentBranch();
-        // Topic boost factors — higher = more likely to surface
-        const topicBoost = TOPIC_BOOST;
-        const results = [];
-        for (const entry of this.entries.values()) {
-            // Filter recent-work by branch (unless branchFilter is "*")
-            if (entry.topic === 'recent-work' && branchFilter !== '*' && entry.branch && entry.branch !== currentBranch) {
-                continue;
-            }
-            // Include tag values as keywords so tagged entries surface in context search
-            const tagKeywordPart = entry.tags ? ` ${entry.tags.join(' ')}` : '';
-            const entryKeywords = extractKeywords(`${entry.title} ${entry.content}${tagKeywordPart}`);
-            const matchedKeywords = [];
-            for (const kw of contextKeywords) {
-                if (entryKeywords.has(kw))
-                    matchedKeywords.push(kw);
+        const allEntries = Array.from(this.entries.values());
+        // Precompute freshness set — keeps ranking functions provably pure (no callbacks)
+        const freshEntryIds = new Set();
+        for (const entry of allEntries) {
+            if (this.isFresh(entry))
+                freshEntryIds.add(entry.id);
+        }
+        const ctx = {
+            currentBranch,
+            branchFilter,
+            topicBoost: TOPIC_BOOST,
+            freshEntryIds,
+            defaultModuleBoost: MODULE_TOPIC_BOOST,
+        };
+        // Keyword ranking (may be empty for stopword-heavy queries)
+        const keywordResults = contextKeywords.size > 0
+            ? keywordRank(allEntries, contextKeywords, minMatch, ctx)
+            : [];
+        // Semantic ranking (only if embedder available and query embeds successfully)
+        const debug = process.env.MEMORY_MCP_DEBUG === '1';
+        let semanticResults = [];
+        if (this.embedder) {
+            const queryResult = await this.embedder.embed(context);
+            if (queryResult.ok) {
+                // In debug mode, get ALL scores (threshold=0) for calibration logging
+                const rawSemanticResults = semanticRank(allEntries, this.vectors, queryResult.vector, debug ? 0 : SEMANTIC_MIN_SIMILARITY, ctx);
+                if (debug) {
+                    for (const r of rawSemanticResults) {
+                        const included = (r.semanticSimilarity ?? 0) >= SEMANTIC_MIN_SIMILARITY;
+                        process.stderr.write(`[memory-mcp:debug] semantic ${(r.semanticSimilarity ?? 0).toFixed(3)} ${r.entry.id} "${r.entry.title}"${included ? '' : ' ← below threshold'}\n`);
+                    }
+                    // Filter to threshold after logging all scores
+                    semanticResults = rawSemanticResults.filter(r => (r.semanticSimilarity ?? 0) >= SEMANTIC_MIN_SIMILARITY);
+                }
+                else {
+                    semanticResults = rawSemanticResults;
+                }
             }
-            if (matchedKeywords.length === 0)
-                continue;
-            // Enforce minimum match threshold
-            const matchRatio = matchedKeywords.length / contextKeywords.size;
-            if (matchRatio < minMatch)
-                continue;
-            // Score = keyword match ratio x confidence x topic boost x reference boost
-            const boost = topicBoost[entry.topic] ?? (entry.topic.startsWith('modules/') ? MODULE_TOPIC_BOOST : 1.0);
-            const freshnessMultiplier = this.isFresh(entry) ? 1.0 : 0.7;
-            // Reference boost: exact class/file name match in references gets a 1.3x multiplier.
-            // Extracts the basename (without extension) from each reference path and stems it,
-            // then checks for overlap with the context keywords.
-            const referenceBoost = entry.references?.some(ref => {
-                const basename = ref.split('/').pop()?.replace(/\.\w+$/, '') ?? ref;
-                return contextKeywords.has(stem(basename.toLowerCase()));
-            }) ? REFERENCE_BOOST_MULTIPLIER : 1.0;
-            // Tag boost: if any tag exactly matches a context keyword, boost the entry
-            const tagBoost = entry.tags?.some(tag => contextKeywords.has(tag))
-                ? TAG_MATCH_BOOST : 1.0;
-            const score = matchRatio * entry.confidence * boost * freshnessMultiplier * referenceBoost * tagBoost;
-            results.push({ entry, score, matchedKeywords });
+            // If embed fails: semanticResults stays empty, keyword results used alone
         }
-        // Always include user entries even if no keyword match (they're always relevant)
+        // Merge keyword + semantic results using max-score strategy
+        const merged = mergeRankings(keywordResults, semanticResults);
+        // Policy: always include user entries even if no keyword/semantic match
+        const results = [...merged];
         for (const entry of this.entries.values()) {
             if (entry.topic === 'user' && !results.find(r => r.entry.id === entry.id)) {
-                results.push({ entry, score: entry.confidence * USER_ALWAYS_INCLUDE_SCORE_FRACTION, matchedKeywords: [] });
+                results.push({
+                    entry,
+                    score: entry.confidence * USER_ALWAYS_INCLUDE_SCORE_FRACTION,
+                    matchedKeywords: [],
+                    source: 'keyword',
+                });
             }
         }
+        // Re-sort after policy injections to maintain score-descending invariant
         return results
             .sort((a, b) => b.score - a.score)
             .slice(0, maxResults);
@@ -914,6 +960,51 @@ export class MarkdownMemoryStore {
             trust: r.entry.trust,
         }));
     }
+    /** Find semantic duplicates by cosine similarity against stored vectors.
+     *  Same-topic only. Returns entries above DEDUP_SEMANTIC_THRESHOLD.
+     *  Returns empty when no embedder or no vectors available. */
+    findSemanticDuplicates(excludeId, topic) {
+        const newVector = this.vectors.get(excludeId);
+        if (!newVector)
+            return [];
+        const related = [];
+        for (const entry of this.entries.values()) {
+            if (entry.id === excludeId)
+                continue;
+            if (entry.topic !== topic)
+                continue;
+            const entryVector = this.vectors.get(entry.id);
+            if (!entryVector)
+                continue;
+            const sim = cosineSimilarity(newVector, entryVector);
+            if (sim > DEDUP_SEMANTIC_THRESHOLD) {
+                related.push({ entry, similarity: sim });
+            }
+        }
+        return related
+            .sort((a, b) => b.similarity - a.similarity)
+            .slice(0, this.behavior.maxDedupSuggestions)
+            .map(r => ({
+            id: r.entry.id,
+            title: r.entry.title,
+            content: r.entry.content,
+            confidence: r.entry.confidence,
+            trust: r.entry.trust,
+        }));
+    }
+    /** Merge keyword-based and semantic-based dedup results, dedup by ID,
+     *  keeping the entry with higher similarity score. */
+    mergeRelatedEntries(keywordDupes, semanticDupes) {
+        const byId = new Map();
+        for (const r of keywordDupes)
+            byId.set(r.id, r);
+        for (const r of semanticDupes) {
+            if (!byId.has(r.id))
+                byId.set(r.id, r);
+            // If already present from keyword dedup, keep whichever — both indicate duplication
+        }
+        return Array.from(byId.values()).slice(0, this.behavior.maxDedupSuggestions);
+    }
     /** Tag frequency across all entries — for vocabulary echo in store responses.
      *  Returns tags sorted by frequency (descending). O(N) over entries. */
     getTagFrequency() {

package/dist/thresholds.d.ts

@@ -14,6 +14,22 @@ export declare const CONFLICT_MIN_CONTENT_CHARS = 50;
 /** Opposition keyword pairs for enhanced conflict detection.
  *  When entries overlap AND use opposing terms, boost the conflict signal. */
 export declare const OPPOSITION_PAIRS: ReadonlyArray<readonly [string, string]>;
+/** Minimum cosine similarity for semantic search results.
+ *  Below this, entries are noise — embedding models produce non-zero similarity
+ *  even for unrelated text.
+ *
+ *  CALIBRATION NOTE: 0.45 is a strict starting point. With nomic-embed-text,
+ *  unrelated text pairs routinely score 0.2-0.4 because the model produces
+ *  non-orthogonal embeddings for any English text. Starting strict and loosening
+ *  with data is safer than starting loose.
+ *
+ *  Use MEMORY_MCP_DEBUG=1 env var to see raw cosine scores for calibration. */
+export declare const SEMANTIC_MIN_SIMILARITY = 0.45;
+/** Minimum cosine similarity for semantic dedup at store time.
+ *  Higher than SEMANTIC_MIN_SIMILARITY because flagging false duplicates is more
+ *  disruptive than missing real ones. Two entries must be quite similar to be
+ *  flagged as potential duplicates. */
+export declare const DEDUP_SEMANTIC_THRESHOLD = 0.8;
 /** Score multiplier when a reference path basename matches the context keywords. */
 export declare const REFERENCE_BOOST_MULTIPLIER = 1.3;
 /** Per-topic scoring boost factors for contextSearch().

package/dist/thresholds.js

@@ -40,6 +40,22 @@ export const OPPOSITION_PAIRS = [
     ['throw', 'return'], // exceptions vs Result types
     ['imperative', 'declarative'],
 ];
+/** Minimum cosine similarity for semantic search results.
+ *  Below this, entries are noise — embedding models produce non-zero similarity
+ *  even for unrelated text.
+ *
+ *  CALIBRATION NOTE: 0.45 is a strict starting point. With nomic-embed-text,
+ *  unrelated text pairs routinely score 0.2-0.4 because the model produces
+ *  non-orthogonal embeddings for any English text. Starting strict and loosening
+ *  with data is safer than starting loose.
+ *
+ *  Use MEMORY_MCP_DEBUG=1 env var to see raw cosine scores for calibration. */
+export const SEMANTIC_MIN_SIMILARITY = 0.45;
+/** Minimum cosine similarity for semantic dedup at store time.
+ *  Higher than SEMANTIC_MIN_SIMILARITY because flagging false duplicates is more
+ *  disruptive than missing real ones. Two entries must be quite similar to be
+ *  flagged as potential duplicates. */
+export const DEDUP_SEMANTIC_THRESHOLD = 0.80;
 /** Score multiplier when a reference path basename matches the context keywords. */
 export const REFERENCE_BOOST_MULTIPLIER = 1.30;
 /** Per-topic scoring boost factors for contextSearch().

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@exaudeus/memory-mcp",
-  "version": "1.6.0",
+  "version": "1.7.0",
   "description": "Codebase memory MCP server - persistent, evolving knowledge for AI coding agents",
   "type": "module",
   "main": "dist/index.js",