@icex-labs/openclaw-memory-engine 5.1.0 → 5.1.1

package/lib/graph.js

@@ -1,9 +1,6 @@
 /**
  * Knowledge Graph: triple store (subject, relation, object).
- * Enables relational queries like "who is George's doctor" or "what treats 荨麻疹".
- *
- * Storage: memory/graph.jsonl — one triple per line.
- * Auto-extraction: pattern-based extraction from archival insert content.
+ * v5.1: strict extraction — only extract triples from clear, short, structured statements.
  */
 
 import { readFileSync, writeFileSync, appendFileSync, existsSync, mkdirSync } from "node:fs";
@@ -12,7 +9,7 @@ import { graphPath } from "./paths.js";
 
 // ─── In-memory cache ───
 
-const cache = new Map(); // ws → { triples: [], loaded: false }
+const cache = new Map();
 
 export function loadGraph(ws) {
   if (cache.has(ws) && cache.get(ws).loaded) return cache.get(ws).triples;
@@ -31,7 +28,6 @@ export function loadGraph(ws) {
 export function addTriple(ws, subject, relation, object, sourceId = null) {
   const triples = loadGraph(ws);
 
-  // Deduplicate: same subject+relation+object
   const exists = triples.some(
     (t) => t.s === subject && t.r === relation && t.o === object,
   );
@@ -61,13 +57,6 @@ export function removeTriple(ws, tripleId) {
   return true;
 }
 
-/**
- * Query the graph from a starting entity, optionally filtering by relation.
- * @param {string} entity - starting node
- * @param {string} [relation] - optional relation filter
- * @param {number} [depth=2] - traversal depth
- * @returns {Array<{ path: string[], triple: object }>}
- */
 export function queryGraph(ws, entity, relation = null, depth = 2) {
   const triples = loadGraph(ws);
   const entityLower = entity.toLowerCase();
@@ -82,16 +71,14 @@ export function queryGraph(ws, entity, relation = null, depth = 2) {
 
     const currentLower = current.toLowerCase();
     for (const t of triples) {
-      // Forward: subject matches
       if (t.s.toLowerCase() === currentLower) {
         if (relation && t.r.toLowerCase() !== relation.toLowerCase()) continue;
-        results.push({ path: [...path, `--${t.r}-->`], node: t.o, triple: t });
+        results.push({ path: [...path, `--${t.r}-->`], node: t.o, triple: { id: t.id, s: t.s, r: t.r, o: t.o } });
         traverse(t.o, currentDepth + 1, [...path, `--${t.r}-->`, t.o]);
       }
-      // Reverse: object matches
       if (t.o.toLowerCase() === currentLower) {
         if (relation && t.r.toLowerCase() !== relation.toLowerCase()) continue;
-        results.push({ path: [...path, `<--${t.r}--`], node: t.s, triple: t });
+        results.push({ path: [...path, `<--${t.r}--`], node: t.s, triple: { id: t.id, s: t.s, r: t.r, o: t.o } });
         traverse(t.s, currentDepth + 1, [...path, `<--${t.r}--`, t.s]);
       }
     }
@@ -101,45 +88,65 @@ export function queryGraph(ws, entity, relation = null, depth = 2) {
   return results;
 }
 
-// ─── Auto-extraction patterns ───
+// ─── Auto-extraction: strict patterns only ───
+
+/**
+ * Validate that a string looks like a proper entity name (not a sentence fragment).
+ * - Must be short (≤25 chars)
+ * - Must not contain markdown, code, or sentence-like patterns
+ * - Must start with a capital letter or CJK character
+ */
+function isValidEntity(s) {
+  if (!s || s.length > 25 || s.length < 2) return false;
+  // Reject markdown, code, URLs, punctuation-heavy strings
+  if (/[`*\[\]{}()|→←⚠#>]/.test(s)) return false;
+  // Reject if it looks like a sentence (has verb-like patterns or too many words)
+  if (s.split(/\s+/).length > 5) return false;
+  // Must start with uppercase, CJK, or known pattern
+  if (!/^[A-Z\u4e00-\u9fff\u3040-\u309f\u30a0-\u30ff\uac00-\ud7af]/.test(s)) return false;
+  return true;
+}
 
 const EXTRACTION_PATTERNS = [
-  // "X's doctor is Y" / "X的医生是Y"
-  { re: /(.+?)(?:'s|的)\s*(?:doctor|医生|主治医生)\s*(?:is|是|为)\s*(.+)/i, r: "has_doctor" },
-  // "X lives in Y" / "X住在Y"
-  { re: /(.+?)\s*(?:lives? in|住在|位于)\s*(.+)/i, r: "lives_in" },
-  // "X works at Y" / "X在Y工作"
-  { re: /(.+?)\s*(?:works? at|在(.+?)工作)/i, r: "works_at" },
-  // "X has condition Y" / disease / 疾病
-  { re: /(.+?)\s*(?:has|有|患有)\s*(?:chronic |慢性)?\s*(.+?(?:症|病|urticaria|condition|disease))/i, r: "has_condition" },
-  // "X treated by/takes Y" / 用药
-  { re: /(.+?)\s*(?:takes?|服用|用药|treated (?:by|with))\s*(.+)/i, r: "treated_by" },
+  // "X's doctor is Y"
+  { re: /^([A-Z]\w+)(?:'s)\s+doctor\s+is\s+(.+)$/i, r: "has_doctor" },
+  // "X's lawyer is Y"
+  { re: /^([A-Z]\w+)(?:'s)\s+(?:lawyer|attorney)\s+is\s+(.+)$/i, r: "has_lawyer" },
+  // "X lives in Y"
+  { re: /^([A-Z]\w+)\s+lives?\s+in\s+([A-Z][\w\s,]+)$/i, r: "lives_in" },
+  // "X works at Y"
+  { re: /^([A-Z]\w+)\s+works?\s+at\s+([A-Z][\w\s&.]+)$/i, r: "works_at" },
+  // "X owns Y" — only match short clear statements
+  { re: /^([A-Z]\w+)\s+owns?\s+(?:a\s+)?([A-Z][\w\s]+)$/i, r: "owns" },
+  // "X drives a Y"
+  { re: /^([A-Z]\w+)\s+drives?\s+(?:a\s+)?([A-Z][\w\s]+)$/i, r: "owns" },
+  // "X has chronic Y" / "X has Y disease/condition"
+  { re: /^([A-Z]\w+)\s+has\s+(?:chronic\s+)?(\w[\w\s]*(?:disease|condition|syndrome|urticaria|diabetes|asthma))$/i, r: "has_condition" },
+  // "X takes Y" (medication)
+  { re: /^([A-Z]\w+)\s+takes?\s+([A-Z][\w\s]+\d+\s*mg)$/i, r: "treated_by" },
+  // "X attends Y"
+  { re: /^([A-Z]\w+)\s+attends?\s+([A-Z][\w\s]+)$/i, r: "attends" },
   // "X's wife/husband is Y"
-  { re: /(.+?)(?:'s|的)\s*(wife|husband|妻子|丈夫|老婆|老公)\s*(?:is|是|为)\s*(.+)/i, r: "spouse", triple: true },
+  { re: /^([A-Z]\w+)(?:'s)\s+(?:wife|husband)\s+is\s+(.+)$/i, r: "spouse" },
   // "X's son/daughter is Y"
-  { re: /(.+?)(?:'s|的)\s*(son|daughter|儿子|女儿)\s*(?:is|是|为)\s*(.+)/i, r: "has_child", triple: true },
-  // "X costs/price Y" / 定价
-  { re: /(.+?)\s*(?:costs?|定价|售价|price[ds]?\s*(?:at)?)\s*\$?([\d,.]+)/i, r: "price" },
-  // "X's lawyer is Y"
-  { re: /(.+?)(?:'s|的)\s*(?:lawyer|律师|attorney)\s*(?:is|是|为)\s*(.+)/i, r: "has_lawyer" },
-  // "X owns Y" / 拥有
-  { re: /(.+?)\s*(?:owns?|拥有|有一辆|drives?)\s*(.+)/i, r: "owns" },
-  // "X studies/attends Y" / 就读
-  { re: /(.+?)\s*(?:attends?|studies? at|就读于?|在(.+?)(?:上学|读书))/i, r: "attends" },
+  { re: /^([A-Z]\w+)(?:'s)\s+(?:son|daughter)\s+is\s+(.+)$/i, r: "has_child" },
 ];
 
 /**
  * Extract triples from a text string.
- * @returns {Array<{ s: string, r: string, o: string }>}
+ * Strict: only matches clean, short statements with proper entity names.
  */
 export function extractTriples(text) {
   const results = [];
+  // Only try extraction on short, clear text (not paragraphs)
+  if (text.length > 150) return results;
+
   for (const pat of EXTRACTION_PATTERNS) {
     const m = text.match(pat.re);
     if (m) {
       const s = (m[1] || "").trim();
-      const o = (pat.triple ? m[3] : m[2] || "").trim();
-      if (s && o && s.length < 50 && o.length < 80) {
+      const o = (m[2] || "").trim();
+      if (isValidEntity(s) && o.length >= 2 && o.length <= 40) {
         results.push({ s, r: pat.r, o });
       }
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@icex-labs/openclaw-memory-engine",
-  "version": "5.1.0",
+  "version": "5.1.1",
   "description": "MemGPT-style hierarchical memory plugin for OpenClaw — core memory block + archival storage with semantic search",
   "type": "module",
   "main": "index.js",