engram-sdk 0.3.2 → 0.3.4

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "engram-sdk",
-  "version": "0.3.2",
+  "version": "0.3.4",
   "description": "Universal memory layer for AI agents. Remember, recall, consolidate.",
   "type": "module",
   "main": "dist/index.js",

package/signal-quality-plan.md

@@ -0,0 +1,23 @@
+# Signal Quality Plan — More Memories, Better Organization
+
+Philosophy: More memories is GOOD. The graph handles scale. The problem isn't quantity — it's classification accuracy in the features that scan broadly (alerts, briefing, consolidation).
+
+## Fix 1: Smarter Status Classification at Intake
+**Problem:** Auto-ingest marks too many things as "pending" — descriptions of future plans in conversation get flagged as commitments even when they're just discussed possibilities.
+**Fix:** Tighten the prompt to distinguish "committed to doing X" from "discussed doing X". Add a post-intake validation in remember() that checks if the content actually represents a real commitment.
+
+## Fix 2: Salience Calibration
+**Problem:** 84% of memories have salience ≥ 0.6, making the signal useless for differentiation.
+**Fix:** Recalibrate extract.ts salience estimation. Most memories should be 0.3-0.5 (normal), with only truly important ones (decisions, preferences, corrections) at 0.7+.
+
+## Fix 3: Briefing as MEMORY.md Replacement
+**Problem:** Briefing dumps a flat list of facts. MEMORY.md is curated with sections and narrative. Agents prefer MEMORY.md.
+**Fix:** Make briefing() cluster memories by entity/topic and present them as organized sections. Add a "what changed since last session" section.
+
+## Fix 4: Alerts Precision
+**Problem:** 238 pending items, most are completed tasks or discussed-but-not-committed plans. False positive rate is too high.
+**Fix:** Add a "fulfilled" detection sweep — if a pending memory's content matches a later completed action, auto-mark it fulfilled. Also add confidence threshold for pending alerts.
+
+## Fix 5: Consolidation Quality Gate
+**Problem:** Consolidation processes all episodes including low-value ones, wasting LLM tokens.
+**Fix:** Filter episodes by salience before consolidation. Only consolidate episodes with salience ≥ 0.3.