lsm / neokai / 26075228515

push

github

Add hybrid semantic memory search (#1947)

* feat: add hybrid agent memory search

Add embedding-backed memory vectors and RRF hybrid ranking so semantic paraphrases can recall Space memories while preserving FTS fallback for pending embeddings.

* fix: harden semantic memory embeddings

Await async query embeddings, ignore stale async writes, persist embedding failures, and avoid truncating vector/paginated search candidates.

* fix: retry failed embedding initialization

Reset failed fastembed initialization, remove unused sync embedding helper, and keep migrated embedding status constraints aligned with bootstrap schema.

* fix: preserve memory search fallback on embedder errors

* fix: harden memory embedding lifecycle

* fix: use named memory key for vector foreign key

* fix: switch memory embeddings to transformers

* fix: pin transformers dependency exactly

* fix: use transformers onnx bge model

* fix: use granite embedding model

* fix: harden embedding backfill coverage

* fix: guard memory embedding backfills

* fix: address memory embedding review feedback

* fix: advance memory embedding backfill past failures

8542 of 12335 branches covered (69.25%)

Branch coverage included in aggregate %.

324 of 337 new or added lines in 9 files covered. (96.14%)

68894 of 80868 relevant lines covered (85.19%)

299.63 hits per line