pomazanbohdan/memory-mcp-1file

Release v0.8.0: The Gemma Migration & Extreme Memory Optimization 🚀

This release marks a fundamental shift in the architecture of memory-mcp, solving the most critical issue reported by users: out-of-memory (OOM) crashes during codebase indexing.

What's New:

• Model Migration (Qwen → Gemma2): We migrated the default embedding model from Qwen3-1.5B to unsloth/embeddinggemma-300m-qat-q4_0-unquantized. This drops the memory footprint of the model from ~3GB down to just ~545MB while maintaining top-tier retrieval performance!
• Zero Configuration: The new Gemma model is fully open. You no longer need a HuggingFace account, HF_TOKEN, or any license agreements to run memory-mcp. It just works out of the box.
• Mimalloc Allocator: Replaced the system allocator with mimalloc. This drastically reduces memory fragmentation (especially on Alpine/Musl) and significantly boosts multi-threaded processing speeds.
• SurrealDB Stability (Throttling): We implemented smart batch-throttling during indexation. The indexer now pauses for 100-150ms after inserting vectors, completely eliminating Transaction write conflict (OCC Retries) inside SurrealDB.
• 2048d Vectors: The model natively generates and searches against 2048-dimensional vectors with last_token_pooling for immense accuracy. The database schema dynamically rebuilds its HNSW indices to accommodate the new dimension.
• Hardware Acceleration: Native release builds now enable x86-64-v3 target optimizations, speeding up the underlying tensor math via AVX2.
• Cleanup: Removed the broken accelerate feature from Cargo to ensure proper compilation on Linux.

Performance:

On a standard system, the container now sits comfortably at ~350MB of RAM usage (down from ~4GB!) during massive codebase indexation, keeping your system fast and responsive.