agent-working-memory 0.7.2 → 0.7.4

package/src/core/salience.ts

@@ -13,7 +13,33 @@
 import type { SalienceFeatures, MemoryClass } from '../types/index.js';
 import type { EngramStore } from '../storage/sqlite.js';
 
-export type SalienceEventType = 'decision' | 'friction' | 'surprise' | 'causal' | 'observation';
+export type SalienceEventType = 'decision' | 'friction' | 'surprise' | 'causal' | 'observation' | 'user_feedback';
+
+/**
+ * Auto-detect user-feedback memories: content that begins with a known user's
+ * name + a feedback verb. These memories represent direct human decisions and
+ * must never be discarded. Examples:
+ *   "Robert verbatim: 'LMS programs-first like CRM'"
+ *   "Katherine said the CEC cycle resets on promotion"
+ *   "Nancy directed Tier 1 has no grace period"
+ *
+ * Why this exists: the BM25 novelty check collapses near-duplicates regardless
+ * of whether the content is a NEW decision or a repeat observation. User
+ * feedback often shares terminology with prior memories ("LMS", "ECP",
+ * "officials") and gets discarded at salience 0.14 (verified in activity log
+ * 2026-05-06T19:08:47). Detecting "Robert said X" → canonical class bypasses
+ * the salience filter entirely.
+ *
+ * Tune the name list as new staff join. Pattern requires word boundary at
+ * start so "Roberta" or "Hannahs" don't match.
+ */
+const USER_FEEDBACK_PATTERN = /^(Robert|Katherine|Catherine|Nancy|Brandy|Brandi|Hannah|Marilyn|Kaylee|Pete|Abby|Tom|Wendy|Sita|Nick|Rob|Joan|Jennifer|Cindy|Jason|Alex|Molly)\s+(said|verbatim|feedback|asked|wants|prefers|requested|requested|directed|decided|confirmed|clarified|chose|specified|explained)\b/i;
+
+/** Returns true if the content looks like direct user feedback that should auto-promote to canonical. */
+export function detectUserFeedback(content: string): boolean {
+  if (typeof content !== 'string' || content.length === 0) return false;
+  return USER_FEEDBACK_PATTERN.test(content.trim());
+}
 
 export interface SalienceInput {
   content: string;
@@ -56,15 +82,28 @@ export function evaluateSalience(
   activeThreshold: number = 0.4,
   stagingThreshold: number = 0.2
 ): SalienceResult {
+  // Auto-detect user feedback before scoring. If content matches the pattern,
+  // force eventType='user_feedback' and memoryClass='canonical'. This bypasses
+  // the BM25 novelty floor that was discarding pivotal user decisions at 0.14.
+  let resolvedEventType: SalienceEventType = input.eventType ?? 'observation';
+  let resolvedMemoryClass: MemoryClass = input.memoryClass ?? 'working';
+  let autoPromoted = false;
+  if (detectUserFeedback(input.content)) {
+    resolvedEventType = 'user_feedback';
+    resolvedMemoryClass = 'canonical';
+    autoPromoted = true;
+  }
+
   const features: SalienceFeatures = {
     surprise: input.surprise ?? 0,
     decisionMade: input.decisionMade ?? false,
     causalDepth: input.causalDepth ?? 0,
     resolutionEffort: input.resolutionEffort ?? 0,
-    eventType: input.eventType ?? 'observation',
+    eventType: resolvedEventType,
   };
 
   const reasonCodes: string[] = [];
+  if (autoPromoted) reasonCodes.push('auto:user_feedback');
 
   // Novelty: 1.0 = completely new info, 0 = exact duplicate exists
   // Default to 0.8 (assume mostly novel) when caller doesn't check
@@ -91,13 +130,14 @@ export function evaluateSalience(
     case 'friction': typeBonus = 0.2; reasonCodes.push('event:friction'); break;
     case 'surprise': typeBonus = 0.25; reasonCodes.push('event:surprise'); break;
     case 'causal': typeBonus = 0.2; reasonCodes.push('event:causal'); break;
+    case 'user_feedback': typeBonus = 0.3; reasonCodes.push('event:user_feedback'); break;
     case 'observation': break;
   }
 
   let score = Math.min(surpriseScore + decisionScore + causalScore + effortScore + noveltyScore + typeBonus, 1.0);
 
   // Memory class overrides
-  const memoryClass = input.memoryClass ?? 'working';
+  const memoryClass = resolvedMemoryClass;
 
   if (memoryClass === 'canonical') {
     // Canonical memories: salience floor of 0.7, never go to staging
@@ -150,17 +190,35 @@ export function computeNovelty(store: EngramStore, agentId: string, concept: str
     // Higher score = stronger match = less novel.
     const topScore = results[0].bm25Score;
 
-    // Penalize exact concept string duplicates — if any result has the same concept,
-    // heavily reduce novelty to prevent hub toxicity from repeated task_end summaries
+    // Quadratic dampening (1 - topScore²) so mid-range matches don't kill novelty.
+    // Old curve was linear (1 - topScore) which floored at 0.1 for almost any match
+    // in a populated DB, killing the salience signal.
+    // Curve comparison (topScore → novelty):
+    //   0.30 → 0.91 (different topic — strong novelty)
+    //   0.60 → 0.64 (loosely related — partial credit)
+    //   0.80 → 0.36 (related but distinct — meaningful signal)
+    //   0.95 → 0.10 (near-dupe — still suppress)
+    const baseNovelty = 1.0 - topScore * topScore;
+
+    // Concept penalty scoped to recent matches only — re-using the same concept
+    // string for a NEW topic months later shouldn't be punished. Penalty was 0.4
+    // (too harsh); now 0.3 and only applies if any matched result is < 30 days old.
     const conceptLower = conceptStr.toLowerCase().trim();
-    const exactConceptMatch = results.some(r => r.engram?.concept?.toLowerCase().trim() === conceptLower);
-    const conceptPenalty = exactConceptMatch ? 0.4 : 0;
-
-    // Continuous novelty: inversely proportional to BM25 similarity
-    // Maps topScore (0..1) → novelty (0.1..0.95) using a smooth curve
-    // Floor at 0.1 (never zero — even duplicates might have new context)
-    // Ceiling at 0.95 (never 1.0 — always a tiny chance of overlap)
-    return Math.max(0.1, Math.min(0.95, 1.0 - topScore - conceptPenalty));
+    const cutoffMs = Date.now() - 30 * 24 * 60 * 60 * 1000;
+    const exactConceptRecent = results.some(r => {
+      if (r.engram?.concept?.toLowerCase().trim() !== conceptLower) return false;
+      const created = r.engram?.createdAt as Date | string | number | undefined;
+      if (!created) return true; // No timestamp — treat as recent (conservative)
+      const createdMs = created instanceof Date
+        ? created.getTime()
+        : typeof created === 'number' ? created : Date.parse(created);
+      return Number.isFinite(createdMs) && createdMs >= cutoffMs;
+    });
+    const conceptPenalty = exactConceptRecent ? 0.3 : 0;
+
+    // Floor lowered to 0.05 (was 0.10) so true duplicates can score near-zero
+    // and clearly stay below stagingThreshold (0.2). Ceiling unchanged.
+    return Math.max(0.05, Math.min(0.95, baseNovelty - conceptPenalty));
   } catch {
     // If BM25 search fails (e.g., FTS not ready), assume novel
     return 0.8;
@@ -206,12 +264,25 @@ export function computeNoveltyWithMatch(
     const top = allResults[0];
     const topScore = top.bm25Score;
 
-    // Concept penalty for exact duplicates
+    // Quadratic dampening — see computeNovelty for curve rationale
+    const baseNovelty = 1.0 - topScore * topScore;
+
+    // Recent-only concept penalty (30d window)
     const conceptLower = conceptStr.toLowerCase().trim();
-    const exactMatch = allResults.some(r => (r.engram as any)?.concept?.toLowerCase().trim() === conceptLower);
-    const conceptPenalty = exactMatch ? 0.4 : 0;
+    const cutoffMs = Date.now() - 30 * 24 * 60 * 60 * 1000;
+    const exactConceptRecent = allResults.some(r => {
+      const eng = r.engram as { concept?: string; createdAt?: Date | string | number };
+      if (eng?.concept?.toLowerCase().trim() !== conceptLower) return false;
+      const created = eng?.createdAt;
+      if (!created) return true;
+      const createdMs = created instanceof Date
+        ? created.getTime()
+        : typeof created === 'number' ? created : Date.parse(created);
+      return Number.isFinite(createdMs) && createdMs >= cutoffMs;
+    });
+    const conceptPenalty = exactConceptRecent ? 0.3 : 0;
 
-    const novelty = Math.max(0.1, Math.min(0.95, 1.0 - topScore - conceptPenalty));
+    const novelty = Math.max(0.05, Math.min(0.95, baseNovelty - conceptPenalty));
     return { novelty, matchedEngramId: top.engram.id, matchScore: topScore };
   } catch {
     return { novelty: 0.8, matchedEngramId: null, matchScore: 0 };

package/src/index.ts

@@ -177,7 +177,7 @@ async function main() {
 
   // Start server
   await app.listen({ port: PORT, host: '0.0.0.0' });
-  console.log(`AgentWorkingMemory v0.7.2 listening on port ${PORT}`);
+  console.log(`AgentWorkingMemory v0.7.4 listening on port ${PORT}`);
 
   // Graceful shutdown
   const shutdown = async () => {

package/src/mcp.ts

@@ -78,7 +78,7 @@ const INCOGNITO = process.env.AWM_INCOGNITO === '1' || process.env.AWM_INCOGNITO
 
 if (INCOGNITO) {
   console.error('AWM: incognito mode — all memory tools disabled, nothing will be recorded');
-  const server = new McpServer({ name: 'agent-working-memory', version: '0.7.2' });
+  const server = new McpServer({ name: 'agent-working-memory', version: '0.7.4' });
   const transport = new StdioServerTransport();
   server.connect(transport).catch(err => {
     console.error('MCP server failed:', err);
@@ -115,7 +115,7 @@ let coordDb: import('better-sqlite3').Database | null = null;
 
 const server = new McpServer({
   name: 'agent-working-memory',
-  version: '0.7.2',
+  version: '0.7.4',
 });
 
 server.registerResource(