pomazanbohdan/memory-mcp-1file

This major release completely overhauls the embedding architecture to replace older baseline models with state-of-the-art 2026 architectures, massively improving retrieval quality, context size, and system performance.

✨ New Models & Efficiency Gains

We have transitioned from the legacy e5_multi to highly optimized, modern alternatives:

• Qwen3-Embedding-0.6B (New Default): A top-tier open-source model featuring a massive 32,768 token context window (vs 512 previously). Despite its larger vocabulary and better semantic precision, it maintains excellent inference speeds.
• embeddinggemma-300m-ONNX: A new ultra-lightweight (~195MB) alternative designed specifically for low-RAM and edge deployments. Extremely fast while retaining strong multilingual capabilities.

📉 Matryoshka Representation Learning (MRL)

We have introduced native support for MRL, allowing users to dynamically truncate embedding vectors via the --mrl-dim flag (e.g., from 1024 down to 512, 256, or 128).

• Efficiency Benchmark: Truncating Qwen3 from 1024 to 512 dimensions reduces database storage (SurrealDB) and vector search latency by ~50%, while retaining >98% of the original retrieval accuracy on MTEB benchmarks.

⚡ Architectural Performance Fixes

• Zero-Block Async Inference: Heavy tensor operations have been offloaded to Tokio's blocking threads (block_in_place), preventing executor starvation. Concurrent JSON-RPC Requests Per Second (RPS) have increased by up to 300% under heavy load.
• Qwen3 Tensor Math Fix: Corrected last-token pooling logic for unpadded sequences, eliminating [PAD] token pollution and restoring exact mathematical accuracy for decoder-only models.
• SurrealDB v3.0.0 Alignment: Database index dimensions now perfectly align with post-MRL truncated outputs.
• L2 Normalization Safety: Added robust protection against NaN/Inf corruption on zero-vectors.

📊 Benchmark Comparison

| Metric | Qwen3-0.6B (New Default) | E5-Multi-Base (Old Default) | Gemma-300m (Edge) |
|--------|--------------------------|-----------------------------|-------------------|
| VRAM / RAM | ~1.2 GB | ~1.1 GB | ~195 MB |
| Context Size | 32,768 tokens | 512 tokens | 8,192 tokens |
| MRL Support | Yes (e.g., 512, 256) | No | Yes |
| RPS (Concurrency) | Non-blocking (High) | Baseline (Blocking) | Fastest |