Expand description
Semantic memory engine with pure-Rust embeddings and git-backed storage.
This crate provides the library core for memory-mcp, an MCP server that
stores and retrieves memories using vector similarity search. Embeddings
are computed on-device via candle (BERT inference) with no C/C++ FFI.
Modules§
- auth
- Token resolution, OAuth device flow, and credential storage backends.
- embedding
- Embedding backends for computing vector representations of text.
- error
- Error types used throughout the crate.
- health
- HTTP health-check handlers (
/readyz). Health reporting infrastructure and HTTP handlers. - index
- HNSW vector index for approximate nearest-neighbour search.
Vector index module — defines the
VectorStoretrait and provides concrete implementations backed by usearch (HNSW) and an in-memory HashMap (for tests). - repo
- Git-backed memory repository — read, write, sync, and diff operations.
- server
- MCP server implementation — tool handlers for the memory protocol.
- types
- Domain types: memories, scopes, metadata, validation, and application state.