WARM vs. Alternatives¶
Hybrid search, graph indexing, and vector-symbolic memory.
WARM (WARM) combines several retrieval and memory paths in one local Rust engine. The paths have different roles; the vector-symbolic layer does not rank ordinary document-search results.
| Need | WARM path | Useful comparison |
|---|---|---|
| Search exact words and phrases | Stemmed BM25 with Boolean and phrase queries | Full-text search engines |
| Search by meaning | Exact cosine over embeddings you supply, fused with lexical results for hybrid search | Embedding-based retrieval systems |
| Search a large vector collection | Graph index with quantized traversal and exact reranking | Approximate nearest-neighbor libraries |
| Store and query structured facts | Vector-symbolic bind, bundle and associative recall, plus a triple store | Symbolic and graph memory systems |
The document search path needs supplied embeddings for semantic matching. Its index can contain source text; enable the encrypted store when confidentiality is required. Hypervector sketches and term counts are not encryption.
The vector-symbolic layer supports composition and associative recall for facts. It is not a substitute for a strong language embedding model when the task is nuanced semantic search over prose. The older sparse-vector path is experimental and trails established ANN libraries; use the graph index for vector search.
For measured wins and losses, see public benchmarks. For the boundary between document retrieval and holographic memory, see what “holographic” means and the architecture map.