claude-memory-layer 1.0.23 → 1.0.24

package/dist/core/index.js

@@ -124,6 +124,7 @@ var ConfigSchema = z.object({
   toolObservation: z.object({
     enabled: z.boolean().default(true),
     excludedTools: z.array(z.string()).default(["TodoWrite", "TodoRead"]),
+    minOutputLength: z.number().default(100),
     maxOutputLength: z.number().default(1e4),
     maxOutputLines: z.number().default(100),
     storeOnlyOnSuccess: z.boolean().default(false)
@@ -2418,6 +2419,21 @@ var SQLiteEventStore = class {
       [id, eventId, sessionId, score, query.slice(0, 100)]
     );
   }
+  /**
+   * Get session IDs that have unevaluated retrievals (measured_at IS NULL).
+   * Excludes the current session. Used to backfill sessions that ended without Stop hook.
+   */
+  async getUnevaluatedSessions(currentSessionId, limit = 5) {
+    await this.initialize();
+    const rows = sqliteAll(
+      this.db,
+      `SELECT DISTINCT session_id FROM memory_helpfulness
+       WHERE measured_at IS NULL AND session_id != ?
+       ORDER BY created_at DESC LIMIT ?`,
+      [currentSessionId, limit]
+    );
+    return rows.map((r) => r.session_id);
+  }
   /**
    * Evaluate helpfulness for all retrievals in a session
    * Called at session end - uses behavioral signals to compute score
@@ -3997,7 +4013,7 @@ var VectorStore = class {
 
 // src/core/embedder.ts
 import { pipeline } from "@huggingface/transformers";
-var Embedder = class {
+var Embedder = class _Embedder {
   pipeline = null;
   modelName;
   activeModelName;
@@ -4028,6 +4044,11 @@ var Embedder = class {
       this.initialized = true;
     }
   }
+  // ~4 chars per token; 512 tokens * 4 = 2048, use 2000 to be safe
+  static MAX_CHARS = 2e3;
+  truncate(text) {
+    return text.length > _Embedder.MAX_CHARS ? text.slice(0, _Embedder.MAX_CHARS) : text;
+  }
   /**
    * Generate embedding for a single text
    */
@@ -4036,10 +4057,11 @@ var Embedder = class {
     if (!this.pipeline) {
       throw new Error("Embedding pipeline not initialized");
     }
-    const output = await this.pipeline(text, {
+    const output = await this.pipeline(this.truncate(text), {
       pooling: "mean",
       normalize: true,
-      truncation: true
+      truncation: true,
+      max_length: 512
     });
     const vector = Array.from(output.data);
     return {
@@ -4061,10 +4083,11 @@ var Embedder = class {
     for (let i = 0; i < texts.length; i += batchSize) {
       const batch = texts.slice(i, i + batchSize);
       for (const text of batch) {
-        const output = await this.pipeline(text, {
+        const output = await this.pipeline(this.truncate(text), {
           pooling: "mean",
           normalize: true,
-          truncation: true
+          truncation: true,
+          max_length: 512
         });
         const vector = Array.from(output.data);
         results.push({