v10.72.0 — Milvus Ranked-Search Parity, Schema Versioning, Belief Store
Special Thanks to two contributors whose work lands in this release:
- @henry201605 — Milvus ranked-mode fix, closing a gap that caused TypeError on every ranked search call against a Milvus backend
- @filhocf — Schema Versioning migration registry + CLI (RFC-1047 §6) and the §2 Belief Store derivation pipeline (RFC-1047 §2)
Fixed
-
fix(milvus): support ranked mode and ranking_weights in search_memories (PR #32, @henry201605)
The ranked search feature (v10.70.0) extended the MemoryStorage.search_memories base signature and the MCP tool handler now always forwards ranking_weights=..., but the Milvus override was never updated. Every memory_search call on a Milvus backend raised TypeError: search_memories() got an unexpected keyword argument 'ranking_weights'. The override now accepts ranking_weights, whitelists ranked mode, over-fetches 3x and applies the shared apply_ranked_rerank multi-signal reranker (semantic + time decay + access frequency + quality), reaching parity with the SQLite-Vec implementation.
Added
-
feat(schema-versioning): migration registry + CLI — issue #11 (PR #13, @filhocf)
Checksum-verified migration registry with run_pending(), _stamp_baseline() for existing databases (probes for schema artifacts to set a correct baseline without re-running already-applied migrations), per-migration transactions, and a new memory schema CLI subcommand (check, migrate, status). The old run_migrations_sync is preserved as a backward-compat wrapper; sqlite_vec.py now routes all schema work through runner.run_pending() directly. 311 tests in tests/test_schema_versioning.py.
-
feat(beliefs): §2 Belief Store — observation-to-belief derivation pipeline (PR #33, @filhocf)
New belief memory subtype. belief.py derivation engine with confidence scoring, contradiction detection, and evidence linking. belief_service.py orchestrator integrating with the existing memory pipeline. SQL migration 012_add_belief_store.sql. 417 tests in tests/test_belief_store.py. Part of the RFC-1047 memory-intelligence track (§0 harvest quality, §1 Observation Store subtypes, §2 Belief Store, §6 schema versioning — now all landed).
Upgrade Notes
No breaking changes. The schema versioning migration runs automatically on startup via the updated sqlite_vec.py init path. Milvus users on ranked mode should upgrade immediately to unblock TypeError failures.
Full changelog: https://codeberg.org/doobidoo/mcp-memory-service/src/branch/main/CHANGELOG.md