@xcidos/genesis-memory-model
v0.1.0-alpha.1
Published
Pinned ONNX embedding model (all-MiniLM-L6-v2) + tokenizer for the Genesis memory server. Platform-independent; shared once across all platforms.
Maintainers
Readme
@xcidos/genesis-memory-model
Platform-independent embedding model for the Genesis MCP memory server. Shipped as a separate package so all eight platform binaries share one copy of the weights instead of bundling them eight times.
Ships two files at the package root:
onnx/model.onnx # ONNX export of the embedder
tokenizer.json # matching fast tokenizerThe launcher (@xcidos/genesis-memory-server) resolves this package's directory and passes
it to the native server as GENESIS_MODEL_DIR.
Model provenance
- Model:
sentence-transformers/all-MiniLM-L6-v2(384-dim, mean pooling) - Pinned revision:
c9745ed1d9f207416be6d2e6f8de32d1f16199bf
The revision is pinned load-bearingly: ONNX exports of the same model differ in output shape and tokenizer behavior, so the pinned commit is what keeps the golden embedding vectors reproducible. Do not change it without regenerating the server's golden vectors and constants.
License
The bundled model and tokenizer are redistributed from all-MiniLM-L6-v2 under
Apache-2.0. This package's metadata is part of the Genesis project
(repository).
