github doobidoo/mcp-memory-service v10.26.8
v10.26.8 - Fix 6 bugs in consolidation, embeddings, and memory types

latest release: v10.26.9
6 hours ago

What's Changed

Six targeted bug fixes across consolidation, embedding cache, and memory type handling.

Fixed

  • [#603] Fix invalid memory_type "learning_note" in learning_session prompt: The create_learning_session prompt handler was emitting memory_type: "learning_note", which is not a valid type in the memory schema. Changed to "learning" so sessions are correctly classified and retrievable by type filter.
  • [#604] Remove memory.touch() call from update_memory_relevance_metadata: update_memory_relevance_metadata was calling memory.touch() as a side-effect, silently overwriting updated_at on every relevance update. Relevance metadata updates no longer corrupt the updated_at timestamp.
  • [#605] Add preserve_timestamps option to update_memories_batch; consolidation callers opt in: update_memories_batch now accepts a preserve_timestamps flag (default False). The consolidation pipeline passes preserve_timestamps=True so merging and compressing memories does not reset their original creation/update times.
  • [#606] Use max(created_at, updated_at) for memory age in _get_memory_age_days: The age calculation previously used only created_at, causing updated memories to decay as if never touched. Now uses max(created_at, updated_at) so a substantive update resets effective age and prevents premature forgetting.
  • [#607] Add dimension fallback when _DIMENSION_CACHE misses on _MODEL_CACHE hit: If the embedding model was cached but the dimension entry was absent, the encoder raised a KeyError. The fix re-derives and caches the dimension from the loaded model without reloading it.
  • [#608] Detect existing DB schema dimension for _HashEmbeddingModel fallback: When the real embedding model cannot load and the _HashEmbeddingModel fallback is used, the code now inspects the existing database schema to determine the vector dimension in use rather than defaulting to a hard-coded value. This prevents dimension mismatch errors when reopening an existing database with the fallback encoder.

Test Coverage

1,340 tests (1,323 existing + 17 new tests covering these fixes).

Full Changelog: https://github.com/doobidoo/mcp-memory-service/blob/main/CHANGELOG.md

Don't miss a new mcp-memory-service release

NewReleases is sending notifications on new releases.