fix(ingest): embed project-scoped writes (#216)

CHANGELOG.md

@@ -54,6 +54,13 @@ the end of this file for historical context.
 - Did not claim the unrelated `longmemory` PyPI project. New Python installs use
   `longmemory-sdk`.
 
+### Fixes
+
+- Embedded project-scoped writes. `longmemory_ingest` with a `project_id`,
+  `longmemory_remember_decision` and connector imports go through
+  `applyImportPlan`, which stored every node without a semantic vector, so
+  those memories never received a vector score in recall (#216).
+
 ## [1.0.0] - 2026-08-31
 
 LongMemory 1.0 is a ground-up architecture rewrite rather than an incremental

benchmarks/src/check.ts

@@ -35,6 +35,7 @@ import { create_named_embedding_provider } from '../../src/core/embeddings/provi
 import { prepare_evidence_query, evidence_support, evidence_adjustment, order_evidence, query_calendar_window, calendar_relevance } from '../../src/core/recall/rerank.js';
 import { associative_recall } from '../../src/core/recall/associative_recall.js';
 import { answer_from_evidence, type evidence_reader_model } from '../../src/answering/evidence_reader.js';
+import { createProjectMemory as create_project_memory } from '../../src/core/project/project_memory.js';
 
 const directory = temp(join(tmpdir(), 'longmemory-bench-check-'));
 const memory = create_memory({ embedding_dimension: 2 });
@@ -451,6 +452,19 @@ try {
             else process.env[key!] = value;
         }
     }
+    {
+        // Project writes and decisions go through applyImportPlan, not ingest(); they must still be embedded.
+        const embedded = create_memory({ embedding_dimension: 4, embedding_provider: { embed: () => [0.5, 0.5, 0.5, 0.5] } });
+        try {
+            const manager = await create_project_memory({ memory: embedded, tenant_id: 'default', project_id: 'check-project', name: 'check' });
+            const id = await manager.ingestProjectEvent('check-project', { kind: 'decision', text: 'Staging credentials rotate every 45 days.' });
+            const node = (await embedded.explain(id)).node;
+            assert.ok(node);
+            assert.equal(node.vectors.semantic?.length, 4);
+        } finally {
+            await embedded.close();
+        }
+    }
     if (!process.argv.includes('--unit-only')) {
         const smoke = await run_benchmark({ providers: ['longmemory'], datasets: ['smoke'], cutoffs: [5], output_dir: join(directory, 'smoke'), resume: false });
         assert.equal(smoke.report.providers[0].failed_questions, 0);

src/core/create_memory.ts

@@ -550,6 +550,12 @@ export function create_memory(config: memory_config = {}): long_memory {
                     throw new Error(`connector node ${node.key} is missing source identity, recorded_at, or provenance`);
                 }
             }
+            // The import transaction is synchronous, so embed before entering it, as ingest() does.
+            const vectors = new Map<(typeof plan.nodes_to_create)[number], number[] | null>();
+            for (const node of plan.nodes_to_create) {
+                const language = detect_language(node.content).language ?? cfg.default_language;
+                vectors.set(node, await embed(cfg.embedding_provider, multilingual_embeddings, embedding_cache, node.content, language, 'document'));
+            }
             return import_transaction().run(() => {
                 const before_nodes = new Set(engine.graph.node_list().map((node) => node.id));
                 const before_edges = new Set(engine.graph.edge_list().map((edge) => edge.id));
@@ -611,6 +617,7 @@ export function create_memory(config: memory_config = {}): long_memory {
                         id: node.id,
                         user_id: cfg.user_id,
                         text: node.content,
+                        vector: vectors.get(node),
                         at: node.recorded_at,
                         observed_at: node.observed_at,
                         valid_from: node.valid_from,